Introduction
Let Ω \Omega Ω be a bounded domain in R n \mathbb{R}^{n} R n , n ≥ 2 n \ge2 n ≥ 2 . Let A = { a i j ( x ) } A=\{a_{ij}(x)\} A = { a ij ( x )} be a measurable, symmetric, positive definite matrix. Consider the elliptic operator in nondivergence form,
L A u = A : D 2 u = : ∑ i , j = 1 n a i j ( x ) D i j u , x ∈ Ω . (1) L_{A}u=A:D^{2}u=:\sum_{i,j=1}^{n}a_{ij}(x)D_{ij}u,\qquad x\in\Omega. \tag*{(1)} L A u = A : D 2 u =: i , j = 1 ∑ n a ij ( x ) D ij u , x ∈ Ω. (1)
We seek the precise range of p p p such that the maximum estimate
sup Ω u ≤ sup ∂ Ω u + + C ( diam Ω ) 2 − n / p ∥ ( A : D 2 u ) − ∥ L p ( Ω ) (2) \sup_{\Omega}u\leq\sup_{\partial\Omega}u^{+}+C(\operatorname{diam}\Omega)^{2-n/p}\left\|(A:D^{2}u)^{-}\right\|_{L^{p}(\Omega)} \tag*{(2)} Ω sup u ≤ ∂ Ω sup u + + C ( diam Ω ) 2 − n / p ( A : D 2 u ) − L p ( Ω ) (2)
holds for functions u u u in the class
u ∈ W l o c 2 , p ( Ω ) ∩ C ( Ω ‾ ) with ( A : D 2 u ) − ∈ L p ( Ω ) , (3) u\in W_{\mathrm{loc}}^{2,p}(\Omega)\cap C(\overline{\Omega})\quad\text{with}\quad(A:D^{2}u)^{-}\in L^{p}(\Omega), \tag*{(3)} u ∈ W loc 2 , p ( Ω ) ∩ C ( Ω ) with ( A : D 2 u ) − ∈ L p ( Ω ) , (3)
where u + = max { u , 0 } u^{+}=\max\{u,0\} u + = max { u , 0 } and u − = max { − u , 0 } u^{-}=\max\{-u,0\} u − = max { − u , 0 } .
When p = n p=n p = n , (2) is the famous Aleksandrov–Bakelman–Pucci maximum principle for linear uniformly elliptic equations; see [1 , 4 , 30 , 14 ] . But the exponent p = n p=n p = n is not optimal for particular operators. For the Laplacian, the optimal range is p > n / 2 p>n/2 p > n /2 . For a general linear uniformly elliptic operator, the optimal exponent depends on an ellipticity constant. A question posed by Pucci is how the optimal exponent depends on the ellipticity condition [29 , 30 ] . For the normalized class
tr A = 1 , A ≥ δ I , 0 < δ ≤ 1 / n , (4) \operatorname{tr}A=1,\qquad A\geq\delta I,\qquad0<\delta\leq1/n, \tag*{(4)} tr A = 1 , A ≥ δ I , 0 < δ ≤ 1/ n , (4)
Pucci [29 ] conjectured that the range is
p > P t r ( n , δ ) : = max { n / 2 , n [ 1 − ( n − 1 ) δ ] } . (5) p>P_{\mathrm{tr}}(n,\delta):=\max\{n/2,n[1-(n-1)\delta]\}. \tag*{(5)} p > P tr ( n , δ ) := max { n /2 , n [ 1 − ( n − 1 ) δ ]} . (5)
The two lower bounds on the right hand side arise from different radial examples which show that the range (5) is sharp for n > 2 n>2 n > 2 [29 ] .
In dimension n = 2 n=2 n = 2 , this conjecture was proved in [3 ] . In higher dimensions, Fabes and Stroock [11 ] obtained an improvement below exponent n n n . Kuo and Trudinger [22 ] proved the maximum estimate under conditions expressed through the dual Gårding cones, giving integer-exponent cases of the conjecture. Trudinger [32 ] refined these results and the planar
2020 Mathematics Subject Classification. 35J15, 35B50.
Key words and phrases. Linear elliptic operators, maximum principle, Pucci conjecture, Hessian equations and integrals, measurable coefficients.
This research was supported by Australian Research Council grant DP230100499 and NSFC No.12225111. estimates. For equations with first-order terms, Trudinger [33 ] obtained the maximum estimate with improved integrability assumptions on the drift under dual-cone conditions. In particular in the uniformly elliptic case, under the normalisation (4) , the drift coefficient b b b need only lie in L q L^q L q for some q > n q>n q > n , as in the Laplacian case. Aleksandrov had originally proved the case p = n p=n p = n with q = n q=n q = n . Krylov [20 , 21 ] extended this to the uniformly elliptic cases where p p p is sufficiently close to n n n , and Jung and Woo [18 ] recently gave an analytic proof of such estimates. See also Dong and Kitano [9 ] for derivative estimates and critical exponents for fully nonlinear equations.
In dimension n = 2 n=2 n = 2 , the Pucci quantification of uniform ellipticity through trace is equivalent to the standard one in terms of the maximum eigenvalue. However, as already remarked in [32 ] , Pucci conjecture alone is not the right one in higher dimensions. In this paper, as well as providing a proof of the full Pucci conjecture, we prove a further sharp result, using the maximum eigenvalue instead of the trace.
First we state the Pucci conjecture result.
Theorem 1.1. Let n ≥ 3 n\geq3 n ≥ 3 and assume (4) . For every p > P t r ( n , δ ) p>P_{\mathrm{tr}}(n,\delta) p > P tr ( n , δ ) , estimate (2) holds for all functions in the class (3) , where the constant depends only on n n n , p p p and δ \delta δ .
For the standard formulation of uniform ellipticity, we consider the the second class of matrices satisfying
λ I ≤ A ( x ) ≤ Λ I , (6) \lambda I\leq A(x)\leq\Lambda I, \tag*{(6)} λ I ≤ A ( x ) ≤ Λ I , (6)
where 0 < λ ≤ Λ < ∞ 0<\lambda\leq\Lambda<\infty 0 < λ ≤ Λ < ∞ . Denote ρ = Λ / λ ≥ 1 \rho=\Lambda/\lambda\geq1 ρ = Λ/ λ ≥ 1 . For 2 ≤ d ≤ n 2\leq d\leq n 2 ≤ d ≤ n and 1 ≤ m < d 1\leq m<d 1 ≤ m < d , define
a d , m = d − m m ρ , λ j = d − j 1 + a d , m , μ j = j a d , m 1 + a d , m . a_{d,m}=\frac{d-m}{m\rho},\qquad\lambda_j=\frac{d-j}{1+a_{d,m}},\qquad\mu_j=\frac{j a_{d,m}}{1+a_{d,m}}. a d , m = m ρ d − m , λ j = 1 + a d , m d − j , μ j = 1 + a d , m j a d , m .
Let K d , m , ρ K_{d,m,\rho} K d , m , ρ be the m × m m\times m m × m matrix given by
K d , m , ρ = ( λ 0 + μ 0 − λ 0 0 ⋯ 0 − μ 1 λ 1 + μ 1 − λ 1 ⋯ 0 0 − μ 2 λ 2 + μ 2 ⋯ 0 ⋮ ⋮ ⋮ ⋱ ⋮ 0 0 ⋯ − μ m − 1 λ m − 1 + μ m − 1 ) (7) K_{d,m,\rho}=
\begin{pmatrix}
\lambda_0+\mu_0 & -\lambda_0 & 0 & \cdots& 0\\
-\mu_1 & \lambda_1+\mu_1 & -\lambda_1 & \cdots& 0\\
0 & -\mu_2 & \lambda_2+\mu_2 & \cdots& 0\\
\vdots& \vdots& \vdots& \ddots& \vdots\\
0 & 0 & \cdots& -\mu_{m-1} & \lambda_{m-1}+\mu_{m-1}
\end{pmatrix} \tag*{(7)} K d , m , ρ = λ 0 + μ 0 − μ 1 0 ⋮ 0 − λ 0 λ 1 + μ 1 − μ 2 ⋮ 0 0 − λ 1 λ 2 + μ 2 ⋮ ⋯ ⋯ ⋯ ⋯ ⋱ − μ m − 1 0 0 0 ⋮ λ m − 1 + μ m − 1 (7)
Denote its smallest eigenvalue by κ d , m ( ρ ) \kappa_{d,m}(\rho) κ d , m ( ρ ) . One has
κ n , 1 ( ρ ) = n ρ ρ + n − 1 . \kappa_{n,1}(\rho)=\frac{n\rho}{\rho+n-1}. κ n , 1 ( ρ ) = ρ + n − 1 n ρ .
Set
P n ( ρ ) = max { n 2 , max 2 ≤ d ≤ n 1 ≤ m < d κ d , m ( ρ ) } (8) P_n(\rho)=\max\left\{\frac{n}{2},\max_{\substack{2\leq d\leq n\\1\leq m<d}}\kappa_{d,m}(\rho)\right\} \tag*{(8)} P n ( ρ ) = max ⎩ ⎨ ⎧ 2 n , 2 ≤ d ≤ n 1 ≤ m < d max κ d , m ( ρ ) ⎭ ⎬ ⎫ (8)
If n = 2 n=2 n = 2 , one can verify directly that P 2 ( ρ ) = P t r ( 2 , 1 / ( 1 + ρ ) ) P_2(\rho)=P_{\mathrm{tr}}(2,1/(1+\rho)) P 2 ( ρ ) = P tr ( 2 , 1/ ( 1 + ρ )) . We will show that
n 2 ≤ P n ( ρ ) < n . (9) \frac{n}{2}\leq P_n(\rho)<n. \tag*{(9)} 2 n ≤ P n ( ρ ) < n . (9)
For the standard formulation of uniform ellipticity, (6) , we have the following sharp maximum estimate.
Theorem 1.2. Let n ≥ 3 n\geq3 n ≥ 3 and assume (6) . For every p > P n ( ρ ) p>P_n(\rho) p > P n ( ρ ) , estimate (2) holds for all functions in the class (3) , where the constant C C C depends only on n n n , p p p , λ , Λ \lambda,\Lambda λ , Λ .
The exponents P t r ( n , δ ) P_{\mathrm{tr}}(n,\delta) P tr ( n , δ ) in Theorem 1.1 and P n ( ρ ) P_n(\rho) P n ( ρ ) in Theorem 1.2 are optimal. When 1 ≤ p ≤ P t r ( n , δ ) 1\leq p\leq P_{\mathrm{tr}}(n,\delta) 1 ≤ p ≤ P tr ( n , δ ) in Theorem 1.1 or 1 < p < P n ( ρ ) 1<p<P_n(\rho) 1 < p < P n ( ρ ) in Theorem 1.2 , estimate (2) does not hold, even for smooth functions in the unit ball, vanishing on the boundary.
A trace-one matrix with A ≥ δ I A \ge\delta I A ≥ δ I also satisfies A ≤ [ 1 − ( n − 1 ) δ ] I A \le[1-(n-1)\delta]I A ≤ [ 1 − ( n − 1 ) δ ] I . Conversely, if I ≤ A ≤ ρ I I \le A \le\rho I I ≤ A ≤ ρ I , then A tr A ≥ I 1 + ( n − 1 ) ρ \frac{A}{\operatorname{tr} A} \ge\frac{I}{1+(n-1)\rho} tr A A ≥ 1 + ( n − 1 ) ρ I . The ranges of the exponent p p p in the above two theorems partially overlap, but neither is completely contained in the other.
We also remark here that from the proof of Theorem 1.1 in [32 ] , we can extend Theorems 1.1 and 1.2 to arbitrary domains Ω \Omega Ω , with the dependence on ( diam Ω ) (\operatorname{diam}\Omega) ( diam Ω ) replaced by that on ∣ Ω ∣ 1 / n |\Omega|^{1/n} ∣Ω ∣ 1/ n , thereby providing an extension to unbounded domains. Corresponding extensions of the local maximum principle, weak and strong Harnack inequalities and Holder estimates of Trudinger, and Krylov-Safonov, as presented for example in [14 ] , also follow. Concerning extensions to operators with first order terms we also have from the proof of the case k ≤ n / 2 k \le n/2 k ≤ n /2 in Theorem 1.1 of [33 ] , that Theorems 1.1 and 1.2 extend to such operators,
L u = A : D 2 u + b ⋅ D u . (10) Lu = A : D^2u + b \cdot Du. \tag*{(10)} Lu = A : D 2 u + b ⋅ D u . (10)
with drift coefficient b b b in L q ( Ω ) L^q(\Omega) L q ( Ω ) , with q = 2 p q=2p q = 2 p , which still does not extend the integer case in [33 ] for p > n / 2 p>n/2 p > n /2 . More generally we need only assume the quantity A − 1 b ⋅ b A^{-1}b\cdot b A − 1 b ⋅ b lies in L p ( Ω ) L^p(\Omega) L p ( Ω ) .
The following is an outline of the proof of Theorems 1.1 and 1.2 .
Choose a noninteger exponent q q q strictly between the proposed threshold and min { p , n } \min\{p,n\} min { p , n } , and put k = ⌊ q ⌋ k=\lfloor q\rfloor k = ⌊ q ⌋ , the largest integer under q q q . The integer k k k selects the Hessian cone Γ k \Gamma_k Γ k , while q q q is the integrability exponent that the estimate must retain. For H ∈ Γ ‾ k H\in\overline{\Gamma}_k H ∈ Γ k , let b b b be its k k k th largest eigenvalue. Subtract the first k + 1 k+1 k + 1 terms of a characteristic polynomial and integrate the remainder against t q − n − 1 d t t^{q-n-1}\,dt t q − n − 1 d t . This defines a continuous function Q n , k , q ( H ) Q_{n,k,q}(H) Q n , k , q ( H ) , homogeneous of degree q q q .
Establish a nonnegative integral for Q Q Q , rather than a pointwise sign. Apply the strong boundary Morse inequalities to t ( ∣ x ∣ 2 − R 2 ) / 2 − w − y ⋅ x t(|x|^2-R^2)/2-w-y\cdot x t ( ∣ x ∣ 2 − R 2 ) /2 − w − y ⋅ x and integrate over y y y . The signed area formula turns the critical-point counts into a truncated determinant integral. Newton’s boundary identities then express the subtracted terms through the normal derivative of w w w on the sphere. Integration in t t t gives ∫ B R Q n , k , q ( D 2 w ) ≥ 0 \int_{B_R}Q_{n,k,q}(D^2w)\ge0 ∫ B R Q n , k , q ( D 2 w ) ≥ 0 for every admissible w w w vanishing on the boundary. The proof also covers a vanishing selected eigenvalue.
Prove two matrix bounds. The first, Q n , k , q ( H ) ≤ C σ k ( H ) q / k Q_{n,k,q}(H)\le C\sigma_k(H)^{q/k} Q n , k , q ( H ) ≤ C σ k ( H ) q / k , holds on the whole cone. The second has the form
Q n , k , q ( H ) ≤ C ( F ( H ) + ) q − c ∣ H ∣ q , Q_{n,k,q}(H)\le C(F(H)^+)^q-c|H|^q, Q n , k , q ( H ) ≤ C ( F ( H ) + ) q − c ∣ H ∣ q ,
where F ( H ) F(H) F ( H ) is the infimum of A : H A:H A : H over the coefficient class. To obtain it, it is enough to prove Q ( H ) < 0 Q(H)<0 Q ( H ) < 0 when H ≠ 0 H\ne0 H = 0 and F ( H ) ≤ 0 F(H)\le0 F ( H ) ≤ 0 ; compactness on ∣ H ∣ = 1 |H|=1 ∣ H ∣ = 1 supplies uniform constants. For fixed trace, separate the smallest eigenvalue. The expression is affine in that eigenvalue, and at the critical value it becomes a weighted sum over deleted pairs. A symmetric compression represents the whole sum. Interlacing and Newton’s inequalities put its roots in the cone needed for a fractional integral sign. Individual summands need not have that sign.
For prescribed eigenvalue bounds, first average the negative eigenvalues. Then maximise over positive eigenvalues with their sum fixed. An extremum argument, including repeated selected eigenvalues, reduces the problem to spectra of the form ( ( − a ) m , 1 d − m ) ((-a)^m,1^{d-m}) (( − a ) m , 1 d − m ) . Zero eigenvalues account for dimensions d < n d<n d < n . The elementary symmetric functions of these spectra are values of a Krawtchouk polynomial. Its first zero is the smallest eigenvalue of K d , m , ρ K_{d,m,\rho} K d , m , ρ . The condition q > P n ( ρ ) q>P_n(\rho) q > P n ( ρ ) puts q q q beyond every relevant first zero and gives the strict matrix sign.
Use an upper obstacle to connect the matrix inequalities to the original function. After subtracting its boundary maximum, assume that u ≤ 0 u\le0 u ≤ 0 on ∂ B R \partial B_R ∂ B R . Solve below
− u + ε -u+\varepsilon − u + ε , with zero boundary values and σ k ( D 2 z ) 1 / k = τ \sigma_k(D^2z)^{1/k}=\tau σ k ( D 2 z ) 1/ k = τ away from contact. At contact, D 2 z ≤ − D 2 u D^2z\le-D^2u D 2 z ≤ − D 2 u , so F ( D 2 z ) ≤ ( A : D 2 u ) − F(D^2z)\le(A:D^2u)^- F ( D 2 z ) ≤ ( A : D 2 u ) − . Use the coercive bound on the contact set and the whole-cone bound elsewhere. Together with the nonnegative integral, they give
∥ σ k ( D 2 z ) 1 / k ∥ L q ≤ C ∥ ( A : D 2 u ) − ∥ L q + C τ R n / q . \left\|\sigma_k(D^2z)^{1/k}\right\|_{L^q}\le C\left\|(A:D^2u)^-\right\|_{L^q}+C\tau R^{n/q}. σ k ( D 2 z ) 1/ k L q ≤ C ( A : D 2 u ) − L q + C τ R n / q .
The Hessian potential estimate bounds the height of z z z by this norm. Its radial integral converges precisely because q > n / 2 q>n/2 q > n /2 . Since u − ε ≤ − z u-\varepsilon\le-z u − ε ≤ − z , sending τ \tau τ and ε \varepsilon ε to zero proves the ball estimate.
(6) Raise q q q to p p p by Hölder’s inequality. Smooth approximation gives the Sobolev estimate on a ball. While we can use the Pucci extremal equation method in [32 ] to extend to arbitrary open bounded sets, we proceed somewhat differently here. On an arbitrary bounded open set, subtract the boundary maximum and a positive constant, take a smooth convex approximation to the positive part, and extend by zero into a containing ball. The resulting function has compact support inside the original set. Its Hessian has an extra positive semidefinite term, which can only improve the one-sided forcing bound.
(7) Prove necessity. Radial solutions with logarithmic cutoffs give failure at n / 2 n/2 n /2 and at the fixed-trace threshold. For the spectral threshold, split diagonal matrices along symmetric rank-one directions, stopping each branch when it has enough negative entries. The same matrix K d , m , ρ K_{d,m,\rho} K d , m , ρ governs the decay of the remaining mass. Below its smallest eigenvalue, the forcing energy tends to zero. At equality, a small change in the negative eigenvalues creates a spectral gap; the accumulated forcing is then of order η p − 1 \eta^{p-1} η p − 1 and still tends to zero. Realise these finite laminates by smooth Hessians and choose smooth coefficients. A convex truncation and rescaling extend the examples from dimension d d d to dimension n n n without losing the critical exponent.
To implement the above argument, we will draw on analytic techniques from works on Hessian equations, including the admissible Hessian theory of [8 , 34 ] , the obstacle regularity theorem of Jiao and Wang [17 ] , and the potential estimate of Phuc and Verbitsky [28 ] ; as well as boundary Morse inequalities in the convention of Laudenbach [23 ] . For the lower estimates, we will use the finite symmetric laminate realization of Astala, Faraco, and Székelyhidi [2 ] , following the convex integration methods of Müller and Šverák [25 ] and the laminate constructions of Faraco [12 ] .
The arrangement of the paper is as follows. Section 2 collects the notation and basic properties of the Hessian cones. Sections 3 and 4 establish the Hessian integral inequality and the matrix estimates used in the obstacle argument of Section 5. Section 6 proves the matrix inequality for coefficients with fixed trace. Sections 7 and 8 treat general ellipticity bounds and identify the spectral threshold P n ( ρ ) P_n(\rho) P n ( ρ ) . These results are combined in Section 9 to prove the upper estimates in Theorems 1.1 and 1.2 . Section 10 gives the radial counterexamples, proving the sharpness of Theorem 1.1 . Sections 11 and 12 construct the spectral counterexamples and extend them to higher dimensions, completing the sharpness proof for Theorem 1.2 .
Acknowledgements
The proofs of both Theorems 1.1 and 1.2 were discovered by ChatGPT more than three weeks ago. The authors have checked and rewritten such input and take responsibility for the full proof.
A different proof of Theorem 1.1 was found, independently by Hung and Wu, (In September, just after ours), and is presented in [16 ] . While reviewing and reorganizing this work, we also learnt that the authors of [24 ] also utilized AI to find a different proof. However, the proof in this paper differs from those, and our Theorem 1.2 is not included in their findings.
More recently the planar case has been treated differently from the approach in [3 ] by Memana, Trudinger and Yung [26 ] leading to the extension of the Pucci conjecture to the non-strict inequality p ≥ P t r ( n , δ ) p \ge P_{\mathrm{tr}}(n,\delta) p ≥ P tr ( n , δ ) for dimension n = 2 n=2 n = 2 and exponent p > 1 p>1 p > 1 . The authors of this paper are grateful to the other authors of these forthcoming papers, for sharing their methods and results
Preliminaries
For H ∈ Sym ( n ) H \in\operatorname{Sym}(n) H ∈ Sym ( n ) , write its eigenvalues in non-decreasing order as λ 1 ( H ) ≤ ⋯ ≤ λ n ( H ) \lambda_1(H) \le\cdots\le\lambda_n(H) λ 1 ( H ) ≤ ⋯ ≤ λ n ( H ) . Write σ j ( H ) \sigma_j(H) σ j ( H ) for the j j j th elementary symmetric function of λ = ( λ 1 , ⋯ , λ n ) \lambda=(\lambda_1,\cdots,\lambda_n) λ = ( λ 1 , ⋯ , λ n ) , with σ 0 = 1 \sigma_0=1 σ 0 = 1 . We use the cones
Γ k = { H : σ j ( H ) > 0 ( 1 ≤ j ≤ k ) } , Γ ‾ k = Γ k ‾ . \Gamma_k=\{H:\sigma_j(H)>0\ (1\le j\le k)\},\qquad\overline{\Gamma}_k=\overline{\Gamma_k}. Γ k = { H : σ j ( H ) > 0 ( 1 ≤ j ≤ k )} , Γ k = Γ k .
and use the same notation for eigenvalue vectors λ \lambda λ . For convenience we denote Γ ‾ 0 = R n \overline{\Gamma}_0=\mathbb{R}^n Γ 0 = R n for eigenvalue vectors and Γ ‾ 0 = Sym ( n ) \overline{\Gamma}_0=\operatorname{Sym}(n) Γ 0 = Sym ( n ) for matrices, and let ∣ H ∣ = ( tr H 2 ) 1 / 2 |H|=(\operatorname{tr}H^2)^{1/2} ∣ H ∣ = ( tr H 2 ) 1/2 be the Frobenius norm. A C 1 , 1 C^{1,1} C 1 , 1 function is k k k -admissible if its Hessian belongs to Γ ‾ k \overline{\Gamma}_k Γ k almost everywhere.
Lemma 2.1. Let 1 ≤ k ≤ n 1\le k\le n 1 ≤ k ≤ n and λ ∈ Γ ‾ k \lambda\in\overline{\Gamma}_k λ ∈ Γ k . Assume that λ 1 ≤ ⋯ ≤ λ n \lambda_1\le\cdots\le\lambda_n λ 1 ≤ ⋯ ≤ λ n . Then
(i) σ k = 0 \sigma_k=0 σ k = 0 on ∂ Γ k \partial\Gamma_k ∂ Γ k and σ k 1 / k ( λ ) \sigma_k^{1/k}(\lambda) σ k 1/ k ( λ ) is concave on Γ k \Gamma_k Γ k .
(ii) ∂ ∂ λ j σ k ( λ ) ≥ 0 ∀ j = 1 , ⋯ , n \frac{\partial}{\partial\lambda_j}\sigma_k(\lambda)\ge0\ \forall j=1,\cdots,n ∂ λ j ∂ σ k ( λ ) ≥ 0 ∀ j = 1 , ⋯ , n .
(iii) Let m = n − k + 1 m=n-k+1 m = n − k + 1 . Then b : = λ m ≥ 0 b:=\lambda_m\ge0 b := λ m ≥ 0 , and
∑ i = 1 m λ i ≥ 0 , − ( m − 1 ) b ≤ λ i ≤ b ( i ≤ m ) . (11) \sum_{i=1}^{m}\lambda_i\ge0,
\qquad
-(m-1)b\le\lambda_i\le b\quad(i\le m). \tag*{(11)} i = 1 ∑ m λ i ≥ 0 , − ( m − 1 ) b ≤ λ i ≤ b ( i ≤ m ) . (11) (iv) We have b > 0 b>0 b > 0 for λ ∈ Γ k \lambda\in\Gamma_k λ ∈ Γ k .
(v) If b = 0 b=0 b = 0 , the matrix is positive semidefinite and has rank at most k − 1 k-1 k − 1 .
Proof. Properties (i) and (ii) are well known. For (iii), delete the largest k − 1 k-1 k − 1 eigenvalues. The remaining list belongs to Γ 1 \Gamma_1 Γ 1 , so its sum is nonnegative. Its largest entry is b b b , which must therefore be nonnegative. For each remaining entry, subtracting the sum of the other m − 1 m-1 m − 1 entries gives the lower bound in (11) . In the open cone the deleted list is in Γ 1 \Gamma_1 Γ 1 , and its sum is positive, so b > 0 b>0 b > 0 . If b = 0 b=0 b = 0 , the lowest m m m entries are nonpositive with nonnegative sum, hence all vanish. The remaining k − 1 k-1 k − 1 entries are nonnegative. □
If a negative entry − a -a − a is deleted from a vector in Γ ‾ k \overline{\Gamma}_k Γ k , the remaining list ξ \xi ξ belongs to Γ k \Gamma_k Γ k , provided it has length at least k k k . Indeed,
σ j ( ξ ) = σ j ( H ) + a σ j − 1 ( ξ ) > 0 ( 1 ≤ j ≤ k ) , \sigma_j(\xi)=\sigma_j(H)+a\sigma_{j-1}(\xi)>0\qquad(1\le j\le k), σ j ( ξ ) = σ j ( H ) + a σ j − 1 ( ξ ) > 0 ( 1 ≤ j ≤ k ) ,
by induction from σ 0 = 1 \sigma_0=1 σ 0 = 1 . The two lower extremal operators are
F δ ( H ) = δ tr H + ( 1 − n δ ) λ 1 ( H ) , F_\delta(H)=\delta\operatorname{tr}H+(1-n\delta)\lambda_1(H), F δ ( H ) = δ tr H + ( 1 − n δ ) λ 1 ( H ) ,
M 1 , ρ ( H ) = ∑ λ i ( H ) > 0 λ i ( H ) + ρ ∑ λ i ( H ) < 0 λ i ( H ) . (12) \mathcal{M}_{1,\rho}(H)=\sum_{\lambda_i(H)>0}\lambda_i(H)+\rho\sum_{\lambda_i(H)<0}\lambda_i(H). \tag*{(12)} M 1 , ρ ( H ) = λ i ( H ) > 0 ∑ λ i ( H ) + ρ λ i ( H ) < 0 ∑ λ i ( H ) . (12)
They are the infima of A : H A:H A : H over { A ≥ δ I : tr A = 1 } \{A\ge\delta I:\operatorname{tr}A=1\} { A ≥ δ I : tr A = 1 } and { I ≤ A ≤ ρ I } \{I\le A\le\rho I\} { I ≤ A ≤ ρ I } , respectively. Both are continuous and positively homogeneous of degree one. Finally, write
B ( s , t ) = ∫ 0 1 r s − 1 ( 1 − r ) t − 1 d r ( s , t > 0 ) . B(s,t)=\int_0^1 r^{s-1}(1-r)^{t-1}\,dr\qquad(s,t>0). B ( s , t ) = ∫ 0 1 r s − 1 ( 1 − r ) t − 1 d r ( s , t > 0 ) .
Boundary Morse inequalities and Hessian integrals
We next obtain an integral inequality without imposing any cone condition. For 0 ≤ r ≤ n − 2 0 \le r \le n-2 0 ≤ r ≤ n − 2 , define
G r ( M ) = { det M if λ r + 1 ( M ) ≥ 0 , 0 otherwise . G_r(M)=
\begin{cases}
\det M & \text{if }\lambda_{r+1}(M)\ge0,\\
0 & \text{otherwise}.
\end{cases} G r ( M ) = { det M 0 if λ r + 1 ( M ) ≥ 0 , otherwise .
Denote by ν \nu ν the unit outer normal.
Lemma 3.1. If U ∈ C 1 , 1 ( B R ‾ ) U \in C^{1,1}(\overline{B_R}) U ∈ C 1 , 1 ( B R ) and U = 0 U=0 U = 0 on ∂ B R \partial B_R ∂ B R , then
( − 1 ) r [ ∫ B R G r ( D 2 U ) d x − 1 n R n − 1 ∫ ∂ B R ( U ν + ) n d S ] ≥ 0 , 0 ≤ r ≤ n − 2. (13) (-1)^r\left[\int_{B_R}G_r(D^2U)\,dx-\frac{1}{nR^{n-1}}\int_{\partial B_R}(U_\nu^+)^n\,dS\right]\ge0,\qquad0\le r\le n-2. \tag*{(13)} ( − 1 ) r [ ∫ B R G r ( D 2 U ) d x − n R n − 1 1 ∫ ∂ B R ( U ν + ) n d S ] ≥ 0 , 0 ≤ r ≤ n − 2. (13) If U U U is uniformly convex, then G r ( D 2 U ) G_r(D^2U) G r ( D 2 U ) is independent of r r r , and one has
∫ B R G r ( D 2 U ) d x = 1 n R n − 1 ∫ ∂ B R ( U ν + ) n d S . \int_{B_R}G_r(D^2U)\,dx=\frac{1}{nR^{n-1}}\int_{\partial B_R}(U_\nu^+)^n\,dS. ∫ B R G r ( D 2 U ) d x = n R n − 1 1 ∫ ∂ B R ( U ν + ) n d S .
Proof. Assume first that U U U is smooth on B R ‾ \overline{B_R} B R . For y ∈ R n y\in\mathbb{R}^n y ∈ R n , set
f y ( x ) = U ( x ) − y ⋅ x . f_y(x)=U(x)-y\cdot x. f y ( x ) = U ( x ) − y ⋅ x .
The critical points of f y f_y f y satisfy D U ( x ) = y DU(x)=y D U ( x ) = y , and their Hessians are D 2 U ( x ) D^2U(x) D 2 U ( x ) . For almost every y ≠ 0 y\ne0 y = 0 , y y y is a regular value of D U DU D U and y ∉ D U ( ∂ B R ) y\notin DU(\partial B_R) y ∈ / D U ( ∂ B R ) . The corresponding critical points are therefore nondegenerate and form a finite set in the interior. Let m j ( y ) m_j(y) m j ( y ) denote the number of these points whose Hessian has exactly j j j negative eigenvalues.
The boundary critical points are particularly simple. Since U = 0 U=0 U = 0 on the sphere,
f y ∣ ∂ B R = − y ⋅ x . f_y|_{\partial B_R}=-y\cdot x. f y ∣ ∂ B R = − y ⋅ x .
This function has two critical points:
x + ( y ) = R y ∣ y ∣ , a minimum of index 0 , x_+(y)=R\frac{y}{|y|},\qquad\text{a minimum of index }0, x + ( y ) = R ∣ y ∣ y , a minimum of index 0 ,
x − ( y ) = − R y ∣ y ∣ , a maximum of index n − 1. x_-(y)=-R\frac{y}{|y|},\qquad\text{a maximum of index }n-1. x − ( y ) = − R ∣ y ∣ y , a maximum of index n − 1.
Here the boundary index is computed from the Hessian in the n − 1 n-1 n − 1 tangential directions. At the minimum,
∂ ν f y ( x + ( y ) ) = U ν ( x + ( y ) ) − ∣ y ∣ . \partial_\nu f_y(x_+(y))=U_\nu(x_+(y))-|y|. ∂ ν f y ( x + ( y )) = U ν ( x + ( y )) − ∣ y ∣.
We now apply the boundary Morse inequalities. Let c j c_j c j count the interior critical points of index j j j , together with the boundary critical points of tangential index j j j for which ∂ ν f y < 0 \partial_\nu f_y<0 ∂ ν f y < 0 . The latter condition means that f y f_y f y increases into the ball. Near such a point, suitable local coordinates ( z , s ) (z,s) ( z , s ) , with s ≥ 0 s\ge0 s ≥ 0 pointing inward, give
f y = f y ( x 0 ) + s + q ( z ) , f_y=f_y(x_0)+s+q(z), f y = f y ( x 0 ) + s + q ( z ) ,
where q q q is a nondegenerate quadratic form of index j j j . These boundary points contribute to changes in the topology of the sublevel sets.
Since the ball is contractible, the polynomial form of the boundary Morse inequalities [23 Theorem A and Corollary A] is
∑ j = 0 n c j t j − 1 = ( 1 + t ) ∑ j ≥ 0 a j t j , a j ≥ 0. \sum_{j=0}^{n}c_jt^j-1=(1+t)\sum_{j\ge0}a_jt^j,\qquad a_j\ge0. j = 0 ∑ n c j t j − 1 = ( 1 + t ) j ≥ 0 ∑ a j t j , a j ≥ 0.
Taking partial alternating sums gives ∑ j = 0 r ( − 1 ) j c j − 1 = ( − 1 ) r a r \sum_{j=0}^{r}(-1)^jc_j-1=(-1)^ra_r ∑ j = 0 r ( − 1 ) j c j − 1 = ( − 1 ) r a r , and hence
( − 1 ) r ( ∑ j = 0 r ( − 1 ) j c j − 1 ) ≥ 0. (14) (-1)^r\left(\sum_{j=0}^{r}(-1)^jc_j-1\right)\ge0. \tag*{(14)} ( − 1 ) r ( j = 0 ∑ r ( − 1 ) j c j − 1 ) ≥ 0. (14)
For r ≤ n − 2 r \le n-2 r ≤ n − 2 , the boundary maximum has index greater than r r r and does not enter this sum. Thus, writing
N 0 ( y ) = 1 { U ν ( R y / ∣ y ∣ ) < ∣ y ∣ } , N_0(y)=\mathbf{1}_{\{U_\nu(Ry/|y|)<|y|\}}, N 0 ( y ) = 1 { U ν ( R y /∣ y ∣ ) < ∣ y ∣ } ,
we have c 0 = m 0 ( y ) + N 0 ( y ) c_0=m_0(y)+N_0(y) c 0 = m 0 ( y ) + N 0 ( y ) and c j = m j ( y ) c_j=m_j(y) c j = m j ( y ) for 1 ≤ j ≤ r 1\le j\le r 1 ≤ j ≤ r . It follows that
( − 1 ) r [ ∑ j = 0 r ( − 1 ) j m j ( y ) − ( 1 − N 0 ( y ) ) ] ≥ 0 (15) (-1)^r\left[\sum_{j=0}^{r}(-1)^j m_j(y)-(1-N_0(y))\right]\ge0 \tag*{(15)} ( − 1 ) r [ j = 0 ∑ r ( − 1 ) j m j ( y ) − ( 1 − N 0 ( y )) ] ≥ 0 (15)
for almost every y y y .
To convert this count into an integral, let E j ⊂ B R E_j\subset B_R E j ⊂ B R be the set where D 2 U D^2U D 2 U is nonsingular and has exactly j j j negative eigenvalues. On this set,
det D 2 U = ( − 1 ) j ∣ det D 2 U ∣ . \det D^2U=(-1)^j|\det D^2U|. det D 2 U = ( − 1 ) j ∣ det D 2 U ∣.
Moreover, λ r + 1 ( D 2 U ) ≥ 0 \lambda_{r+1}(D^2U)\ge0 λ r + 1 ( D 2 U ) ≥ 0 precisely when j ≤ r j\le r j ≤ r . Since G r G_r G r vanishes at singular Hessians, the area formula gives
∫ B R G r ( D 2 U ) d x = ∑ j = 0 r ( − 1 ) j ∫ E j ∣ det D 2 U ∣ d x = ∫ R n ∑ j = 0 r ( − 1 ) j m j ( y ) d y . (16) \begin{aligned}
\int_{B_R}G_r(D^2U)\,dx
&=\sum_{j=0}^{r}(-1)^j\int_{E_j}|\det D^2U|\,dx\\
&=\int_{\mathbb{R}^n}\sum_{j=0}^{r}(-1)^j m_j(y)\,dy. \tag*{(16)}
\end{aligned} ∫ B R G r ( D 2 U ) d x = j = 0 ∑ r ( − 1 ) j ∫ E j ∣ det D 2 U ∣ d x = ∫ R n j = 0 ∑ r ( − 1 ) j m j ( y ) d y . (16)
For the other term, write y = s θ y=s\theta y = s θ , with s > 0 s>0 s > 0 and θ ∈ S n − 1 \theta\in S^{n-1} θ ∈ S n − 1 . By the definition of N 0 N_0 N 0 ,
1 − N 0 ( s θ ) = 1 { s ≤ U ν ( R θ ) } . 1-N_0(s\theta)=\mathbf{1}_{\{s\le U_\nu(R\theta)\}}. 1 − N 0 ( s θ ) = 1 { s ≤ U ν ( R θ )} .
Integration in polar coordinates therefore yields
∫ R n ( 1 − N 0 ( y ) ) d y = ∫ S n − 1 ∫ 0 U ν ( R θ ) + s n − 1 d s d θ = 1 n R n − 1 ∫ ∂ B R ( U ν + ) n d S . \begin{aligned}
\int_{\mathbb{R}^n}(1-N_0(y))\,dy
&=\int_{S^{n-1}}\int_{0}^{U_\nu(R\theta)^+}s^{n-1}\,ds\,d\theta\\
&=\frac{1}{nR^{n-1}}\int_{\partial B_R}(U_\nu^+)^n\,dS.
\end{aligned} ∫ R n ( 1 − N 0 ( y )) d y = ∫ S n − 1 ∫ 0 U ν ( R θ ) + s n − 1 d s d θ = n R n − 1 1 ∫ ∂ B R ( U ν + ) n d S .
Integrating (15) and using these identities proves (13) for smooth U U U .
Finally, let U ∈ C 1 , 1 ( B R ‾ ) U\in C^{1,1}(\overline{B_R}) U ∈ C 1 , 1 ( B R ) . Fix s > n s>n s > n , choose g i ∈ C ∞ ( B R ‾ ) g_i\in C^\infty(\overline{B_R}) g i ∈ C ∞ ( B R ) converging to Δ U \Delta U Δ U in L s ( B R ) L^s(B_R) L s ( B R ) , and solve
Δ U i = g i in B R , U i = 0 on ∂ B R . \Delta U_i=g_i\quad\text{in }B_R,\qquad U_i=0\quad\text{on }\partial B_R. Δ U i = g i in B R , U i = 0 on ∂ B R .
The solutions are smooth on the closed ball. The Dirichlet estimate gives
∥ U i − U ∥ W 2 , s ( B R ) ≤ C ∥ g i − Δ U ∥ L s ( B R ) ⟶ 0. \|U_i-U\|_{W^{2,s}(B_R)}\le C\|g_i-\Delta U\|_{L^s(B_R)}\longrightarrow0. ∥ U i − U ∥ W 2 , s ( B R ) ≤ C ∥ g i − Δ U ∥ L s ( B R ) ⟶ 0.
Consequently, U i → U U_i\to U U i → U in C 1 ( B R ‾ ) C^1(\overline{B_R}) C 1 ( B R ) and D 2 U i → D 2 U D^2U_i\to D^2U D 2 U i → D 2 U in measure. The function G r G_r G r is continuous: its indicator changes only where λ r + 1 = 0 \lambda_{r+1}=0 λ r + 1 = 0 , and the determinant vanishes there. Thus G r ( D 2 U i ) → G r ( D 2 U ) G_r(D^2U_i)\to G_r(D^2U) G r ( D 2 U i ) → G r ( D 2 U ) in measure. Also, ∣ G r ( H ) ∣ ≤ C n ∣ H ∣ n |G_r(H)|\le C_n|H|^n ∣ G r ( H ) ∣ ≤ C n ∣ H ∣ n , so for every measurable E ⊂ B R E\subset B_R E ⊂ B R ,
∫ E ∣ G r ( D 2 U i ) ∣ d x ≤ C n ∥ D 2 U i ∥ L s ( B R ) n ∣ E ∣ 1 − n / s ≤ C ∣ E ∣ 1 − n / s . \int_E|G_r(D^2U_i)|\,dx\le C_n\|D^2U_i\|_{L^s(B_R)}^n|E|^{1-n/s}\le C|E|^{1-n/s}. ∫ E ∣ G r ( D 2 U i ) ∣ d x ≤ C n ∥ D 2 U i ∥ L s ( B R ) n ∣ E ∣ 1 − n / s ≤ C ∣ E ∣ 1 − n / s .
Since s > n s>n s > n , this gives uniform integrability. Together with convergence in measure, it implies convergence of the bulk integrals. The boundary integrals converge by the uniform convergence of D U i DU_i D U i . Passing to the limit proves (13) . □
Fix 1 ≤ k ≤ n − 1 1\le k\le n-1 1 ≤ k ≤ n − 1 and k < p < k + 1 k<p<k+1 k < p < k + 1 . For H ∈ Γ ‾ k H\in\overline{\Gamma}_k H ∈ Γ k put b = λ n − k + 1 ( H ) b=\lambda_{n-k+1}(H) b = λ n − k + 1 ( H ) and define
Q n , k , p ( H ) = { ( − 1 ) k ∑ j = 0 n ( − 1 ) j σ j ( H ) b p − j p − j , b > 0 , 0 , b = 0. (17) Q_{n,k,p}(H)=
\begin{cases}
(-1)^k\sum_{j=0}^{n}\frac{(-1)^j\sigma_j(H)b^{p-j}}{p-j}, & b>0,\\
0, & b=0.
\end{cases} \tag*{(17)} Q n , k , p ( H ) = { ( − 1 ) k ∑ j = 0 n p − j ( − 1 ) j σ j ( H ) b p − j , 0 , b > 0 , b = 0. (17)
The selected eigenvalue b b b is the k k k th largest eigenvalue.
Proposition 3.2. The function Q n , k , p Q_{n,k,p} Q n , k , p is continuous on Γ ‾ k \overline{\Gamma}_k Γ k , is positively homogeneous of degree p p p , and satisfies ∣ Q n , k , p ( H ) ∣ ≤ C ∣ H ∣ p \lvert Q_{n,k,p}(H)\rvert\le C\lvert H\rvert^p ∣ Q n , k , p ( H )∣ ≤ C ∣ H ∣ p . For every zero-boundary k k k -admissible w ∈ C 1 , 1 ( B R ‾ ) w \in C^{1,1}(\overline{B_R}) w ∈ C 1 , 1 ( B R ) ,
∫ B R Q n , k , p ( D 2 w ) d x ≥ ( n − 1 ) ! ∣ ∏ j = 0 n ( p − j ) ∣ R 1 − p ∫ ∂ B R w ν p d S ≥ 0. (18) \int_{B_R} Q_{n,k,p}(D^2w)\,dx \ge\frac{(n-1)!}{\left\lvert\prod_{j=0}^{n}(p-j)\right\rvert} R^{1-p}\int_{\partial B_R}w_\nu^p\,dS \ge0. \tag*{(18)} ∫ B R Q n , k , p ( D 2 w ) d x ≥ ∏ j = 0 n ( p − j ) ( n − 1 )! R 1 − p ∫ ∂ B R w ν p d S ≥ 0. (18)
Proof. By Lemma 2.1 , the lowest n − k + 1 n-k+1 n − k + 1 eigenvalues have absolute value at most C n b C_n b C n b . A product of j ≥ k j \ge k j ≥ k eigenvalues contains at least j − k + 1 j-k+1 j − k + 1 of these entries. Consequently
∣ σ j ( H ) ∣ ≤ C n b j − k + 1 ∣ H ∣ k − 1 ( j ≥ k ) , \lvert\sigma_j(H)\rvert\le C_n b^{j-k+1}\lvert H\rvert^{k-1}\qquad(j\ge k), ∣ σ j ( H )∣ ≤ C n b j − k + 1 ∣ H ∣ k − 1 ( j ≥ k ) ,
∣ σ j ( H ) ∣ ≤ C n ∣ H ∣ j ( j < k ) . \lvert\sigma_j(H)\rvert\le C_n\lvert H\rvert^j\qquad(j<k). ∣ σ j ( H )∣ ≤ C n ∣ H ∣ j ( j < k ) .
After multiplication by b p − j b^{p-j} b p − j , these inequalities give the asserted growth and continuity at b = 0 b=0 b = 0 , because p − k + 1 > 0 p-k+1>0 p − k + 1 > 0 . Homogeneity and continuity where b > 0 b>0 b > 0 follow directly from the definition.
To prove the integral inequality, subtract the first k + 1 k+1 k + 1 terms of the characteristic polynomial and set
L t ( H ) = G k − 1 ( t I − H ) − ∑ j = 0 k ( − 1 ) j σ j ( H ) t n − j . L_t(H)=G_{k-1}(tI-H)-\sum_{j=0}^{k}(-1)^j\sigma_j(H)t^{n-j}. L t ( H ) = G k − 1 ( t I − H ) − j = 0 ∑ k ( − 1 ) j σ j ( H ) t n − j .
The indicator in G k − 1 ( t I − H ) G_{k-1}(tI-H) G k − 1 ( t I − H ) switches at t = b t=b t = b . Therefore
L t ( H ) = { − ∑ j = 0 k ( − 1 ) j σ j ( H ) t n − j , 0 < t < b , ∑ j = k + 1 n ( − 1 ) j σ j ( H ) t n − j , t > b . L_t(H)=
\begin{cases}
-\sum_{j=0}^{k}(-1)^j\sigma_j(H)t^{n-j}, & 0<t<b,\\[6pt]
\sum_{j=k+1}^{n}(-1)^j\sigma_j(H)t^{n-j}, & t>b.
\end{cases} L t ( H ) = ⎩ ⎨ ⎧ − ∑ j = 0 k ( − 1 ) j σ j ( H ) t n − j , ∑ j = k + 1 n ( − 1 ) j σ j ( H ) t n − j , 0 < t < b , t > b .
Integrating each monomial gives
( − 1 ) k − 1 ∫ 0 ∞ t p − n − 1 L t ( H ) d t = Q n , k , p ( H ) . (19) (-1)^{k-1}\int_0^\infty t^{p-n-1}L_t(H)\,dt=Q_{n,k,p}(H). \tag*{(19)} ( − 1 ) k − 1 ∫ 0 ∞ t p − n − 1 L t ( H ) d t = Q n , k , p ( H ) . (19)
The first integral converges at zero since p > k p>k p > k , and the second at infinity since p < k + 1 p<k+1 p < k + 1 . The same eigenvalue bounds as above show that its absolute integral is at most C ∣ H ∣ p C\lvert H\rvert^p C ∣ H ∣ p . If b = 0 b=0 b = 0 , the rank assertion in Lemma 2.1 gives L t ( H ) = 0 L_t(H)=0 L t ( H ) = 0 .
Let q = w ν q=w_\nu q = w ν on the sphere. Since Δ w ≥ 0 \Delta w\ge0 Δ w ≥ 0 and w = 0 w=0 w = 0 on the boundary, w ≤ 0 w\le0 w ≤ 0 and q ≥ 0 q\ge0 q ≥ 0 . We use the Newton tensor boundary identities [31 ]
∫ B R σ j ( D 2 w ) d x = 1 n ( n j ) R 1 − j ∫ ∂ B R q j d S , 1 ≤ j ≤ n . (20) \int_{B_R}\sigma_j(D^2w)\,dx=\frac{1}{n}\binom{n}{j}R^{1-j}\int_{\partial B_R}q^j\,dS,\qquad1\le j\le n. \tag*{(20)} ∫ B R σ j ( D 2 w ) d x = n 1 ( j n ) R 1 − j ∫ ∂ B R q j d S , 1 ≤ j ≤ n . (20)
For smooth w w w , the Newton tensor
T j − 1 a b = ∂ σ j ( D 2 w ) / ∂ w a b T_{j-1}^{ab}=\partial\sigma_j(D^2w)/\partial w_{ab} T j − 1 ab = ∂ σ j ( D 2 w ) / ∂ w ab
is divergence free, and
j σ j ( D 2 w ) = ∂ a ( T j − 1 a b w b ) . j\sigma_j(D^2w)=\partial_a(T_{j-1}^{ab}w_b). j σ j ( D 2 w ) = ∂ a ( T j − 1 ab w b ) .
On the sphere, D w = q ν Dw=q\nu D w = q ν and the tangential Hessian is ( q / R ) I (q/R)I ( q / R ) I . Hence
T j − 1 ν ν = ( n − 1 j − 1 ) ( q / R ) j − 1 , T_{j-1}^{\nu\nu}=\binom{n-1}{j-1}(q/R)^{j-1}, T j − 1 ν ν = ( j − 1 n − 1 ) ( q / R ) j − 1 ,
which gives (20) . The Poisson approximation used in Lemma 3.1 extends this identity to C 1 , 1 C^{1,1} C 1 , 1 functions. For j = 0 j=0 j = 0 , the corresponding formula is simply ∣ B R ∣ = R ∣ ∂ B R ∣ / n \lvert B_R\rvert=R\lvert\partial B_R\rvert/n ∣ B R ∣ = R ∣ ∂ B R ∣ / n .
Apply Lemma 3.1 with r = k − 1 r=k-1 r = k − 1 to U = t ( ∣ x ∣ 2 − R 2 ) / 2 − w U=t(\lvert x\rvert^2-R^2)/2-w U = t (∣ x ∣ 2 − R 2 ) /2 − w . Subtracting the integrals in (20) gives
( − 1 ) k − 1 [ ∫ B R L t ( D 2 w ) d x − ∫ ∂ B R Ψ t ( q ) d S ] ≥ 0 , (-1)^{k-1}\left[\int_{B_R}L_t(D^2w)\,dx-\int_{\partial B_R}\Psi_t(q)\,dS\right]\ge0, ( − 1 ) k − 1 [ ∫ B R L t ( D 2 w ) d x − ∫ ∂ B R Ψ t ( q ) d S ] ≥ 0 ,
where
Ψ t ( q ) = R n [ ( t − q / R ) + n − ∑ j = 0 k ( − 1 ) j ( n j ) ( q / R ) j t n − j ] . \Psi_t(q)=\frac{R}{n}\left[(t-q/R)_+^n-\sum_{j=0}^{k}(-1)^j\binom{n}{j}(q/R)^jt^{n-j}\right]. Ψ t ( q ) = n R [ ( t − q / R ) + n − j = 0 ∑ k ( − 1 ) j ( j n ) ( q / R ) j t n − j ] .
Multiply by t p − n − 1 t^{p-n-1} t p − n − 1 and integrate over t > 0 t>0 t > 0 . For q > 0 q>0 q > 0 , split the scalar integral at t = q / R t=q/R t = q / R and integrate its monomials, exactly as in (19) . The identity
∑ j = 0 n ( − 1 ) j ( n j ) p − j = ( − 1 ) n n ! ∏ j = 0 n ( p − j ) \sum_{j=0}^{n}\frac{(-1)^j\binom{n}{j}}{p-j}=\frac{(-1)^n n!}{\prod_{j=0}^{n}(p-j)} j = 0 ∑ n p − j ( − 1 ) j ( j n ) = ∏ j = 0 n ( p − j ) ( − 1 ) n n !
then gives
( − 1 ) k − 1 ∫ 0 ∞ t p − n − 1 Ψ t ( q ) d t = ( n − 1 ) ! R 1 − p q p ∏ j = 0 n ( p − j ) . (-1)^{k-1}\int_{0}^{\infty}t^{p-n-1}\Psi_t(q)\,\mathrm{d}t=\frac{(n-1)!R^{1-p}q^p}{\prod_{j=0}^{n}(p-j)}. ( − 1 ) k − 1 ∫ 0 ∞ t p − n − 1 Ψ t ( q ) d t = ∏ j = 0 n ( p − j ) ( n − 1 )! R 1 − p q p .
For q = 0 q=0 q = 0 , both sides are zero. Absolute convergence, with the preceding growth bounds and boundedness of D 2 w D^2w D 2 w and q q q , justifies Fubini’s theorem and proves (18) . □
Matrix estimates
The obstacle argument needs an upper bound on the whole Hessian cone. We first establish two polynomial signs, then use them to prove the bound and a fractional sign inequality used in both coefficient classes.
Lemma 4.1. Let h 1 ≥ ⋯ ≥ h d h_1\ge\cdots\ge h_d h 1 ≥ ⋯ ≥ h d belong to Γ ‾ l \overline{\Gamma}_l Γ l , where 1 ≤ l ≤ d 1\le l\le d 1 ≤ l ≤ d , and suppose c = h l > 0 c=h_l>0 c = h l > 0 . Set
D h ( t ) = ∏ i = 1 d ( 1 − t h i ) , D_h(t)=\prod_{i=1}^{d}(1-th_i), D h ( t ) = i = 1 ∏ d ( 1 − t h i ) , E l ( h , t ) = D h ( t ) − ∑ j = 0 l − 1 ( − 1 ) j σ j ( h ) t j , E_l(h,t)=D_h(t)-\sum_{j=0}^{l-1}(-1)^j\sigma_j(h)t^j, E l ( h , t ) = D h ( t ) − j = 0 ∑ l − 1 ( − 1 ) j σ j ( h ) t j , T l ( h , t ) = ( − 1 ) l − 1 ∑ j = 0 l − 1 ( − 1 ) j σ j ( h ) t j . T_l(h,t)=(-1)^{l-1}\sum_{j=0}^{l-1}(-1)^j\sigma_j(h)t^j. T l ( h , t ) = ( − 1 ) l − 1 j = 0 ∑ l − 1 ( − 1 ) j σ j ( h ) t j . Then
( − 1 ) l E l ( h , t ) ≥ 0 ( 0 ≤ t ≤ c − 1 ) , (-1)^lE_l(h,t)\ge0\qquad(0\le t\le c^{-1}), ( − 1 ) l E l ( h , t ) ≥ 0 ( 0 ≤ t ≤ c − 1 ) , T l ( h , t ) ≥ ( c t − 1 ) l − 1 ( t ≥ c − 1 ) . T_l(h,t)\ge(ct-1)^{l-1}\qquad(t\ge c^{-1}). T l ( h , t ) ≥ ( c t − 1 ) l − 1 ( t ≥ c − 1 ) . Proof. Put L = h 1 ≥ c L=h_1\ge c L = h 1 ≥ c . For 0 ≤ s < L − 1 0\le s<L^{-1} 0 ≤ s < L − 1 , every factor in D h ( s ) D_h(s) D h ( s ) is positive and
h i 1 − s h i − h i = s h i 2 1 − s h i ≥ 0. \frac{h_i}{1-sh_i}-h_i=\frac{sh_i^2}{1-sh_i}\ge0. 1 − s h i h i − h i = 1 − s h i s h i 2 ≥ 0. The vector with entries h i / ( 1 − s h i ) h_i/(1-sh_i) h i / ( 1 − s h i ) is therefore in Γ ‾ l \overline{\Gamma}_l Γ l . Differentiating the product gives
( − 1 ) l D h ( l ) ( s ) = l ! D h ( s ) σ l ( h 1 1 − s h 1 , … , h d 1 − s h d ) ≥ 0. (-1)^lD_h^{(l)}(s)=l!D_h(s)\sigma_l\left(\frac{h_1}{1-sh_1},\ldots,\frac{h_d}{1-sh_d}\right)\ge0. ( − 1 ) l D h ( l ) ( s ) = l ! D h ( s ) σ l ( 1 − s h 1 h 1 , … , 1 − s h d h d ) ≥ 0. Taylor’s integral formula proves the first sign up to t = L − 1 t=L^{-1} t = L − 1 . This proves its full range when l = 1 l=1 l = 1 .
For l ≥ 2 l\ge2 l ≥ 2 , delete L L L and call the remaining vector h ′ h' h ′ . It belongs to Γ ‾ l − 1 \overline{\Gamma}_{l-1} Γ l − 1 and its ( l − 1 ) (l-1) ( l − 1 ) st largest entry is c c c . Expanding D h ( t ) = ( 1 − L t ) D h ′ ( t ) D_h(t)=(1-Lt)D_{h'}(t) D h ( t ) = ( 1 − L t ) D h ′ ( t ) gives
( − 1 ) l E l ( h , t ) = ( L t − 1 ) ( − 1 ) l − 1 E l − 1 ( h ′ , t ) + σ l − 1 ( h ′ ) t l − 1 , (-1)^lE_l(h,t)=(Lt-1)(-1)^{l-1}E_{l-1}(h',t)+\sigma_{l-1}(h')t^{l-1}, ( − 1 ) l E l ( h , t ) = ( L t − 1 ) ( − 1 ) l − 1 E l − 1 ( h ′ , t ) + σ l − 1 ( h ′ ) t l − 1 , T l ( h , t ) = ( L t − 1 ) T l − 1 ( h ′ , t ) + σ l − 1 ( h ′ ) t l − 1 . T_l(h,t)=(Lt-1)T_{l-1}(h',t)+\sigma_{l-1}(h')t^{l-1}. T l ( h , t ) = ( L t − 1 ) T l − 1 ( h ′ , t ) + σ l − 1 ( h ′ ) t l − 1 .
The first identity and induction extend the first sign over L − 1 ≤ t ≤ c − 1 L^{-1} \le t \le c^{-1} L − 1 ≤ t ≤ c − 1 . For the second sign, start with T 1 = 1 T_1=1 T 1 = 1 . If t ≥ c − 1 t \ge c^{-1} t ≥ c − 1 , then L t − 1 ≥ c t − 1 ≥ 0 Lt-1 \ge ct-1 \ge0 L t − 1 ≥ c t − 1 ≥ 0 . The second identity, deletion, and induction give T l ( h , t ) ≥ ( c t − 1 ) l − 1 T_l(h,t) \ge(ct-1)^{l-1} T l ( h , t ) ≥ ( c t − 1 ) l − 1 . □ \square □
Proposition 4.2. For 1 ≤ k ≤ n − 1 1 \le k \le n-1 1 ≤ k ≤ n − 1 and k < p < k + 1 k < p < k+1 k < p < k + 1 , there is a constant C = C ( n , k , p ) C=C(n,k,p) C = C ( n , k , p ) such that
Q n , k , p ( H ) ≤ C σ k ( H ) p / k ( H ∈ Γ ‾ k ) . Q_{n,k,p}(H) \le C\sigma_k(H)^{p/k} \qquad(H \in\overline{\Gamma}_k). Q n , k , p ( H ) ≤ C σ k ( H ) p / k ( H ∈ Γ k ) . Proof. We first show a strict sign when σ k ( H ) = 0 \sigma_k(H)=0 σ k ( H ) = 0 and b = λ n − k + 1 ( H ) > 0 b=\lambda_{n-k+1}(H)>0 b = λ n − k + 1 ( H ) > 0 . Use E k E_k E k and T k T_k T k from Lemma 4.1 , applied to the eigenvalues of H H H . Since σ k ( H ) = 0 \sigma_k(H)=0 σ k ( H ) = 0 , the remainder E k ( H , t ) E_k(H,t) E k ( H , t ) is O ( t k + 1 ) O(t^{k+1}) O ( t k + 1 ) at zero. Termwise integration in (17) gives
Q n , k , p ( H ) = ( − 1 ) k ∑ j = 0 k − 1 ( − 1 ) j σ j ( H ) b p − j p − j − ∫ 0 1 / b t − p − 1 ( − 1 ) k E k ( H , t ) d t = − ∫ 1 / b ∞ t − p − 1 T k ( H , t ) d t − ∫ 0 1 / b t − p − 1 ( − 1 ) k E k ( H , t ) d t ≤ − ∫ 1 / b ∞ t − p − 1 ( b t − 1 ) k − 1 d t = − b p B ( p − k + 1 , k ) < 0. \begin{aligned}
Q_{n,k,p}(H) &= (-1)^k\sum_{j=0}^{k-1}\frac{(-1)^j\sigma_j(H)b^{p-j}}{p-j}-\int_0^{1/b}t^{-p-1}(-1)^kE_k(H,t)\,\mathrm{d}t\\
&=-\int_{1/b}^{\infty}t^{-p-1}T_k(H,t)\,\mathrm{d}t-\int_0^{1/b}t^{-p-1}(-1)^kE_k(H,t)\,\mathrm{d}t\\
&\le-\int_{1/b}^{\infty}t^{-p-1}(bt-1)^{k-1}\,\mathrm{d}t=-b^pB(p-k+1,k)<0.
\end{aligned} Q n , k , p ( H ) = ( − 1 ) k j = 0 ∑ k − 1 p − j ( − 1 ) j σ j ( H ) b p − j − ∫ 0 1/ b t − p − 1 ( − 1 ) k E k ( H , t ) d t = − ∫ 1/ b ∞ t − p − 1 T k ( H , t ) d t − ∫ 0 1/ b t − p − 1 ( − 1 ) k E k ( H , t ) d t ≤ − ∫ 1/ b ∞ t − p − 1 ( b t − 1 ) k − 1 d t = − b p B ( p − k + 1 , k ) < 0. Here B ( a , b ) = ∫ 0 1 s a − 1 ( 1 − s ) b − 1 d s B(a,b)=\int_0^1s^{a-1}(1-s)^{b-1}\,\mathrm{d}s B ( a , b ) = ∫ 0 1 s a − 1 ( 1 − s ) b − 1 d s . Both integrals converge because k < p < k + 1 k<p<k+1 k < p < k + 1 .
Normalize b = 1 b=1 b = 1 and compactify the largest k − 1 k-1 k − 1 eigenvalues by taking their reciprocals. Write
H = diag ( a 1 , … , a m , L 1 , … , L k − 1 ) , H=\operatorname{diag}(a_1,\ldots,a_m,L_1,\ldots,L_{k-1}), H = diag ( a 1 , … , a m , L 1 , … , L k − 1 ) , where m = n − k + 1 m=n-k+1 m = n − k + 1 , a m = 1 a_m=1 a m = 1 , a i ≤ 1 a_i\le1 a i ≤ 1 , L i ≥ 1 L_i\ge1 L i ≥ 1 . Lemma 2.1 gives a i ≥ − ( m − 1 ) a_i\ge-(m-1) a i ≥ − ( m − 1 ) . Set
P = ∏ i = 1 k − 1 L i ≥ 1 , r i = L i − 1 , S = σ k ( H ) P , F = Q n , k , p ( H ) P . P=\prod_{i=1}^{k-1}L_i\ge1,\qquad r_i=L_i^{-1},\qquad S=\frac{\sigma_k(H)}{P},\qquad F=\frac{Q_{n,k,p}(H)}{P}. P = i = 1 ∏ k − 1 L i ≥ 1 , r i = L i − 1 , S = P σ k ( H ) , F = P Q n , k , p ( H ) . For k = 1 k=1 k = 1 , the product is understood to be 1. Both S S S and F F F are polynomials in ( a , r ) (a,r) ( a , r ) ; division by P P P replaces each omitted large eigenvalue by its reciprocal. Thus they extend continuously to the compact closure K K K of the admissible normalized vectors in the box a i ∈ [ − ( m − 1 ) , 1 ] a_i\in[-(m-1),1] a i ∈ [ − ( m − 1 ) , 1 ] , r i ∈ [ 0 , 1 ] r_i\in[0,1] r i ∈ [ 0 , 1 ] .
We claim F < 0 F<0 F < 0 at every point of K K K where S = 0 S=0 S = 0 . If all r i > 0 r_i>0 r i > 0 , this is the strict sign just proved. Otherwise, suppose exactly ℓ \ell ℓ reciprocals vanish and delete the corresponding diverging eigenvalues from an approximating sequence. Let H ∗ H_* H ∗ be the limiting finite vector and P ∗ P_* P ∗ the product of its remaining L i L_i L i entries. Deletion gives H ∗ ∈ Γ ‾ k − ℓ H_*\in\overline{\Gamma}_{k-\ell} H ∗ ∈ Γ k − ℓ , and expansion by the diverging eigenvalues gives
S = σ k − ℓ ( H ∗ ) P ∗ , F = Q n − ℓ , k − ℓ , p − ℓ ( H ∗ ) P ∗ . S=\frac{\sigma_{k-\ell}(H_*)}{P_*},\qquad F=\frac{Q_{n-\ell,k-\ell,p-\ell}(H_*)}{P_*}. S = P ∗ σ k − ℓ ( H ∗ ) , F = P ∗ Q n − ℓ , k − ℓ , p − ℓ ( H ∗ ) . For example, the terms surviving in σ j / P \sigma_j/P σ j / P are precisely those containing all ℓ \ell ℓ diverging entries; their remaining degree is j − ℓ j-\ell j − ℓ . This also explains the shift p ↦ p − ℓ p\mapsto p-\ell p ↦ p − ℓ in Q Q Q . The selected eigenvalue of H ∗ H_* H ∗ remains 1 and k − ℓ < p − ℓ < k − ℓ + 1 k-\ell<p-\ell<k-\ell+1 k − ℓ < p − ℓ < k − ℓ + 1 . The strict sign in the smaller dimension therefore proves the claim.
Since S ≥ 0 S\ge0 S ≥ 0 on K K K , compactness gives a number δ > 0 \delta>0 δ > 0 such that F > 0 F>0 F > 0 implies S ≥ δ S\ge\delta S ≥ δ . Indeed, otherwise a sequence with F > 0 F>0 F > 0 and S → 0 S\to0 S → 0 would have a limit in K K K with S = 0 S=0 S = 0 and F ≥ 0 F\ge0 F ≥ 0 , contrary to the strict sign just proved. Also F ≤ C F\le C F ≤ C on K K K . At a point where Q n , k , p ( H ) > 0 Q_{n,k,p}(H)>0 Q n , k , p ( H ) > 0 ,
Q n , k , p ( H ) = P F ≤ C P ≤ C P p / k ≤ C δ − p / k ( P S ) p / k = C δ − p / k σ k ( H ) p / k . Q_{n,k,p}(H)=PF\le CP\le CP^{p/k}\le C\delta^{-p/k}(PS)^{p/k}=C\delta^{-p/k}\sigma_k(H)^{p/k}. Q n , k , p ( H ) = P F ≤ C P ≤ C P p / k ≤ C δ − p / k ( P S ) p / k = C δ − p / k σ k ( H ) p / k . If Q ≤ 0 Q\le0 Q ≤ 0 , the assertion is immediate. Homogeneity removes the normalization b = 1 b=1 b = 1 , and continuity gives the assertion at b = 0 b=0 b = 0 . □ \square □
For a finite list r r r and a noninteger real number q q q , set
V q ( r ) = ∑ j = 0 # r ( − 1 ) j σ j ( r ) q − j . (21) V_q(r)=\sum_{j=0}^{\#r}\frac{(-1)^j\sigma_j(r)}{q-j}. \tag*{(21)} V q ( r ) = j = 0 ∑ # r q − j ( − 1 ) j σ j ( r ) . (21)
For the empty list, V q ( ∅ ) = 1 / q V_q(\varnothing)=1/q V q ( ∅ ) = 1/ q . In the next lemma lists are ordered decreasingly.
Lemma 4.3. Let r 1 ≥ ⋯ ≥ r d r_1\ge\cdots\ge r_d r 1 ≥ ⋯ ≥ r d belong to Γ ‾ l \overline{\Gamma}_l Γ l , where 1 ≤ l ≤ d 1\le l\le d 1 ≤ l ≤ d , and let c = r l > 0 c=r_l>0 c = r l > 0 . If l − 1 < q < l l-1<q<l l − 1 < q < l , then
( − 1 ) l + 1 c q V q ( r / c ) ≥ c q B ( q − l + 1 , l ) > 0. (-1)^{l+1}c^qV_q(r/c)\ge c^qB(q-l+1,l)>0. ( − 1 ) l + 1 c q V q ( r / c ) ≥ c q B ( q − l + 1 , l ) > 0. The same lower bound holds with c q V q ( r / c ) c^qV_q(r/c) c q V q ( r / c ) replaced by b q V q ( r / b ) b^qV_q(r/b) b q V q ( r / b ) whenever b > 0 b>0 b > 0 and
r l + 1 ≤ b ≤ r l − 1 . (22) r_{l+1}\le b\le r_{l-1}. \tag*{(22)} r l + 1 ≤ b ≤ r l − 1 . (22) An endpoint whose index is outside 1 , … , d 1,\ldots,d 1 , … , d imposes no restriction.
Proof. Use E l E_l E l and T l T_l T l from Lemma 4.1. Since E l ( r , t ) = O ( t l ) E_l(r,t)=O(t^l) E l ( r , t ) = O ( t l ) at zero and l − 1 < q < l l-1<q<l l − 1 < q < l , termwise integration gives
( − 1 ) l c q V q ( r / c ) = − ∫ 1 / c ∞ t − q − 1 T l ( r , t ) d t − ∫ 0 1 / c t − q − 1 ( − 1 ) l E l ( r , t ) d t (-1)^lc^qV_q(r/c)=-\int_{1/c}^{\infty}t^{-q-1}T_l(r,t)\,dt-\int_0^{1/c}t^{-q-1}(-1)^lE_l(r,t)\,dt ( − 1 ) l c q V q ( r / c ) = − ∫ 1/ c ∞ t − q − 1 T l ( r , t ) d t − ∫ 0 1/ c t − q − 1 ( − 1 ) l E l ( r , t ) d t ≤ − ∫ 1 / c ∞ t − q − 1 ( c t − 1 ) l − 1 d t = − c q B ( q − l + 1 , l ) . \le-\int_{1/c}^{\infty}t^{-q-1}(ct-1)^{l-1}\,dt=-c^qB(q-l+1,l). ≤ − ∫ 1/ c ∞ t − q − 1 ( c t − 1 ) l − 1 d t = − c q B ( q − l + 1 , l ) . This proves the sign at b = c b=c b = c .
To see why it persists throughout (22) , define F ( b ) = ( − 1 ) l + 1 b q V q ( r / b ) F(b)=(-1)^{l+1}b^qV_q(r/b) F ( b ) = ( − 1 ) l + 1 b q V q ( r / b ) for b > 0 b>0 b > 0 . Differentiation gives
F ′ ( b ) = ( − 1 ) l + 1 b q − d − 1 ∏ i = 1 d ( b − r i ) . F'(b)=(-1)^{l+1}b^{q-d-1}\prod_{i=1}^{d}(b-r_i). F ′ ( b ) = ( − 1 ) l + 1 b q − d − 1 i = 1 ∏ d ( b − r i ) . On the positive part of [ r l + 1 , r l ] [r_{l+1},r_l] [ r l + 1 , r l ] this derivative is nonpositive, because the product has l l l negative factors. On [ r l , r l − 1 ] [r_l,r_{l-1}] [ r l , r l − 1 ] it is nonnegative, because there are l − 1 l-1 l − 1 such factors. Thus F F F attains its minimum over the stated interval at c = r l c=r_l c = r l . Repeated roots give the same weak signs by continuity.
The Hessian obstacle argument
We use the following obstacle and potential estimates.
First, if h ∈ C ∞ ( B R ‾ ) h\in C^\infty(\overline{B_R}) h ∈ C ∞ ( B R ) is positive on the boundary and τ > 0 \tau>0 τ > 0 , there is a k k k -admissible function z τ ∈ C 1 , 1 ( B R ‾ ) z_\tau\in C^{1,1}(\overline{B_R}) z τ ∈ C 1 , 1 ( B R ) such that
z τ = 0 on ∂ B R , z τ ≤ h , σ k ( D 2 z τ ) 1 / k ≥ τ a.e. in B R , σ k ( D 2 z τ ) 1 / k = τ in { z τ < h } . (23) \begin{aligned}
z_\tau&=0\text{ on }\partial B_R,\qquad z_\tau\le h,\\
\sigma_k(D^2z_\tau)^{1/k}&\ge\tau\text{ a.e. in }B_R,\\
\sigma_k(D^2z_\tau)^{1/k}&=\tau\text{ in }\{z_\tau<h\}. \tag*{(23)}
\end{aligned} z τ σ k ( D 2 z τ ) 1/ k σ k ( D 2 z τ ) 1/ k = 0 on ∂ B R , z τ ≤ h , ≥ τ a.e. in B R , = τ in { z τ < h } . (23)
This follows from [17 Theorems 1.1 and 1.2] , with κ = 0 \kappa=0 κ = 0 , boundary value zero, and right-hand side τ \tau τ . The admissible subsolution is C ( ∣ x ∣ 2 − R 2 ) C(|x|^2-R^2) C ( ∣ x ∣ 2 − R 2 ) for sufficiently large C C C , and the auxiliary strictly convex function is ∣ x ∣ 2 |x|^2 ∣ x ∣ 2 . The operator σ k 1 / k \sigma_k^{1/k} σ k 1/ k is among the examples covered by their theorem. The C 1 , 1 C^{1,1} C 1 , 1 viscosity inequalities hold almost everywhere by testing with second-order Taylor polynomials plus ϵ ∣ x − x 0 ∣ 2 \epsilon|x-x_0|^2 ϵ ∣ x − x 0 ∣ 2 and letting ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 . On the noncontact set the solution is classical.
Second, for a zero-boundary k k k -admissible z ∈ C 1 , 1 ( B R ‾ ) z\in C^{1,1}(\overline{B_R}) z ∈ C 1 , 1 ( B R ) and p > max { k , n / 2 } p>\max\{k,n/2\} p > max { k , n /2 } ,
∥ z ∥ L ∞ ( B R ) ≤ C n , k , p R 2 − n / p ∥ σ k ( D 2 z ) 1 / k ∥ L p ( B R ) . (24) \|z\|_{L^\infty(B_R)}\le C_{n,k,p}R^{2-n/p}\left\|\sigma_k(D^2z)^{1/k}\right\|_{L^p(B_R)}. \tag*{(24)} ∥ z ∥ L ∞ ( B R ) ≤ C n , k , p R 2 − n / p σ k ( D 2 z ) 1/ k L p ( B R ) . (24)
Put F = σ k ( D 2 z ) 1 / k F=\sigma_k(D^2z)^{1/k} F = σ k ( D 2 z ) 1/ k . Subharmonicity gives z ≤ 0 z\leq0 z ≤ 0 . We first identify the Hessian measure used in the potential estimate. On a relatively compact subdomain, mollify z z z to obtain z ε z_\varepsilon z ε . The convexity of the elementary symmetric matrix cone implies D 2 z ε ∈ Γ k D^2z_\varepsilon\in\Gamma_k D 2 z ε ∈ Γ k . Moreover, z ε → z z_\varepsilon\to z z ε → z locally uniformly, so z z z is also k k k -subharmonic in the viscosity sense; see [34 Section 2] . The Hessians are locally uniformly bounded and converge in every finite L s L^s L s space. Hence σ k ( D 2 z ε ) → F k \sigma_k(D^2z_\varepsilon)\to F^k σ k ( D 2 z ε ) → F k in L l o c 1 L^1_{\mathrm{loc}} L loc 1 . Weak continuity of Hessian measures, in [28 Theorem 7.1] , gives μ k [ z ] = F k d x \mu_k[z]=F^k\,dx μ k [ z ] = F k d x . Applying [28 Theorem 7.4] to − z -z − z on the ball now gives
− z ( x ) ≤ C n , k ∫ 0 4 R ( ∫ B r ( x ) ∩ B R F k d y r n − 2 k ) 1 / k d r r . -z(x)\leq C_{n,k}\int_0^{4R}\left(\frac{\int_{B_r(x)\cap B_R}F^k\,dy}{r^{n-2k}}\right)^{1/k}\frac{dr}{r}. − z ( x ) ≤ C n , k ∫ 0 4 R ( r n − 2 k ∫ B r ( x ) ∩ B R F k d y ) 1/ k r d r .
The ball is uniformly ( k − 1 ) (k-1) ( k − 1 ) -convex, and F k ∈ L p / k ( B R ) F^k\in L^{p/k}(B_R) F k ∈ L p / k ( B R ) with p / k > max { 1 , n / ( 2 k ) } p/k>\max\{1,n/(2k)\} p / k > max { 1 , n / ( 2 k )} , so the hypotheses of the potential estimate hold. Hölder’s inequality bounds the expression in parentheses to the power 1 / k 1/k 1/ k by C r 2 − n / p ∥ F ∥ L p ( B R ) Cr^{2-n/p}\lVert F\rVert_{L^p(B_R)} C r 2 − n / p ∥ F ∥ L p ( B R ) . The integral of r 2 − n / p d r / r r^{2-n/p}\,dr/r r 2 − n / p d r / r over ( 0 , 4 R ) (0,4R) ( 0 , 4 R ) is finite since p > n / 2 p>n/2 p > n /2 , and equals a constant times R 2 − n / p R^{2-n/p} R 2 − n / p . This proves (24) .
Lemma 5.1. Let 1 ≤ k ≤ n − 1 1\leq k\leq n-1 1 ≤ k ≤ n − 1 , k < p < k + 1 k<p<k+1 k < p < k + 1 , and p > n / 2 p>n/2 p > n /2 . Let A \mathcal{A} A be a class of symmetric matrices uniformly bounded above and below by positive multiples of the identity and set F A ( H ) = inf A ∈ A A : H F_{\mathcal{A}}(H)=\inf_{A\in\mathcal{A}}A:H F A ( H ) = inf A ∈ A A : H . Suppose a continuous function Q Q Q on Γ k \Gamma_k Γ k has nonnegative integral on every zero-boundary k k k -admissible C 1 , 1 C^{1,1} C 1 , 1 function on a ball, and there are positive constants C 0 , C 1 , c 1 C_0,C_1,c_1 C 0 , C 1 , c 1 such that
Q ( H ) ≤ C 0 σ k ( H ) p / k , Q(H)\leq C_0\sigma_k(H)^{p/k}, Q ( H ) ≤ C 0 σ k ( H ) p / k , Q ( H ) ≤ C 1 ( F A ( H ) + ) p − c 1 σ k ( H ) p / k . Q(H)\leq C_1\bigl(F_{\mathcal{A}}(H)^+\bigr)^p-c_1\sigma_k(H)^{p/k}. Q ( H ) ≤ C 1 ( F A ( H ) + ) p − c 1 σ k ( H ) p / k . Then, for every measurable A ( x ) ∈ A A(x)\in\mathcal{A} A ( x ) ∈ A and every u ∈ W 2 , p ( B R ) u\in W^{2,p}(B_R) u ∈ W 2 , p ( B R ) ,
sup B R u ≤ sup ∂ B R u + + C R 2 − n / p ∥ ( A : D 2 u ) − ∥ L p ( B R ) . \sup_{B_R}u\leq\sup_{\partial B_R}u^+ + CR^{2-n/p}\left\|(A:D^2u)^-\right\|_{L^p(B_R)}. B R sup u ≤ ∂ B R sup u + + C R 2 − n / p ( A : D 2 u ) − L p ( B R ) .
Proof. First assume u u u is smooth on the closed ball and u ≤ 0 u\leq0 u ≤ 0 on its boundary. Let ε > 0 \varepsilon>0 ε > 0 , set h = − u + ε h=-u+\varepsilon h = − u + ε , and choose z τ z_\tau z τ from (23) . Denote its contact set by K = { z τ = h } K=\{z_\tau=h\} K = { z τ = h } and write f = ( A : D 2 u ) − f=(A:D^2u)^- f = ( A : D 2 u ) − . At almost every point of K K K , the function h − z τ h-z_\tau h − z τ has a local minimum zero and is twice differentiable. Hence
D 2 z τ ≤ D 2 h = − D 2 u . D^2z_\tau\leq D^2h=-D^2u. D 2 z τ ≤ D 2 h = − D 2 u .
Since A A A is positive definite,
F A ( D 2 z τ ) ≤ A : D 2 z τ ≤ − A : D 2 u ≤ f . F_{\mathcal{A}}(D^2z_\tau)\leq A:D^2z_\tau\leq-A:D^2u\leq f. F A ( D 2 z τ ) ≤ A : D 2 z τ ≤ − A : D 2 u ≤ f .
Use the second inequality of (5.3) on K K K . On B R ∖ K B_R\setminus K B R ∖ K , use the first inequality and σ k ( D 2 z τ ) 1 / k = τ \sigma_k(D^2z_\tau)^{1/k}=\tau σ k ( D 2 z τ ) 1/ k = τ . Nonnegativity of the integral of Q Q Q then gives
0 ≤ ∫ B R Q ( D 2 z τ ) d x ≤ C 1 ∫ K f p d x − c 1 ∫ K σ k ( D 2 z τ ) p / k d x + C 0 τ p ∣ B R ∖ K ∣ . 0\leq\int_{B_R}Q(D^2z_\tau)\,dx\leq C_1\int_K f^p\,dx-c_1\int_K\sigma_k(D^2z_\tau)^{p/k}\,dx+C_0\tau^p|B_R\setminus K|. 0 ≤ ∫ B R Q ( D 2 z τ ) d x ≤ C 1 ∫ K f p d x − c 1 ∫ K σ k ( D 2 z τ ) p / k d x + C 0 τ p ∣ B R ∖ K ∣.
Adding the known contribution off K K K , we conclude that
∫ B R σ k ( D 2 z τ ) p / k d x ≤ C ∫ B R f p d x + C τ p ∣ B R ∣ . \int_{B_R}\sigma_k(D^2z_\tau)^{p/k}\,dx\leq C\int_{B_R}f^p\,dx+C\tau^p|B_R|. ∫ B R σ k ( D 2 z τ ) p / k d x ≤ C ∫ B R f p d x + C τ p ∣ B R ∣.
Since z τ ≤ − u + ε z_\tau\leq-u+\varepsilon z τ ≤ − u + ε , we have u − ε ≤ − z τ u-\varepsilon\leq-z_\tau u − ε ≤ − z τ . Applying (24) yields
sup B R u − ε ≤ C R 2 − n / p ∥ f ∥ L p ( B R ) + C τ R 2 . \sup_{B_R}u-\varepsilon\leq CR^{2-n/p}\lVert f\rVert_{L^p(B_R)}+C\tau R^2. B R sup u − ε ≤ C R 2 − n / p ∥ f ∥ L p ( B R ) + C τ R 2 .
Let τ ↓ 0 \tau\downarrow0 τ ↓ 0 and then ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . Subtracting sup ∂ B R u + \sup_{\partial B_R}u^+ sup ∂ B R u + gives (5.4) for arbitrary smooth u u u .
Finally approximate a function in W 2 , p ( B R ) W^{2,p}(B_R) W 2 , p ( B R ) by functions smooth on the closed ball. Since p > n / 2 p>n/2 p > n /2 , the so A : D 2 u j → A : D 2 u A : D^2u_j \to A : D^2u A : D 2 u j → A : D 2 u in L p L^p L p . The negative-part map is Lipschitz, and therefore the right-hand sides converge as well. This proves (5.4). □
The matrix inequality for fixed trace
For the fixed-trace class, write A = δ I + B A = \delta I + B A = δ I + B , where B ≥ 0 B \ge0 B ≥ 0 and tr B = 1 − n δ \operatorname{tr} B = 1 - n\delta tr B = 1 − n δ . Minimizing B : H B : H B : H places its trace in an eigendirection of the least eigenvalue of H H H , giving F δ F_\delta F δ in (12) . We will show that Q n , k , p ( H ) < 0 Q_{n,k,p}(H) < 0 Q n , k , p ( H ) < 0 whenever H ≠ 0 H \ne0 H = 0 and F δ ( H ) ≤ 0 F_\delta(H) \le0 F δ ( H ) ≤ 0 , provided p > n [ 1 − ( n − 1 ) δ ] p > n[1 - (n - 1)\delta] p > n [ 1 − ( n − 1 ) δ ] . Throughout this section, lists other than the original eigenvalues λ i ( H ) \lambda_i(H) λ i ( H ) are ordered decreasingly. We use V q V_q V q from (21) .
The following polynomial allows us to treat a weighted sum of deleted pairs at once; the eigenvalue separation used in its proof is the standard min–max principle, as in [15 ] . Its individual summands need not have the same sign.
Lemma 6.1. Let N ≥ 2 N \ge2 N ≥ 2 , let μ 1 ≥ ⋯ ≥ μ N \mu_1 \ge\cdots\ge\mu_N μ 1 ≥ ⋯ ≥ μ N be a nonconstant real vector, and set S = ∑ i μ i S = \sum_i \mu_i S = ∑ i μ i . Define
P ( t ) = ∏ i = 1 N ( t − μ i ) , P(t) = \prod_{i=1}^{N}(t-\mu_i), P ( t ) = i = 1 ∏ N ( t − μ i ) , R ( t ) = ∑ i < j ( μ i − μ j ) 2 ∏ h ≠ i , j ( t − μ h ) . R(t) = \sum_{i<j}(\mu_i-\mu_j)^2\prod_{h\ne i,j}(t-\mu_h). R ( t ) = i < j ∑ ( μ i − μ j ) 2 h = i , j ∏ ( t − μ h ) . Then
R ( t ) = ( N t − S ) P ′ ( t ) − N 2 P ( t ) = D 2 ∏ i = 1 N − 2 ( t − r i ) , R(t) = (Nt-S)P'(t)-N^2P(t)=D_2\prod_{i=1}^{N-2}(t-r_i), R ( t ) = ( N t − S ) P ′ ( t ) − N 2 P ( t ) = D 2 i = 1 ∏ N − 2 ( t − r i ) , where D 2 = N ∑ i μ i 2 − S 2 > 0 D_2 = N\sum_i\mu_i^2-S^2>0 D 2 = N ∑ i μ i 2 − S 2 > 0 , the roots are real and, in decreasing order, satisfy
μ i + 2 ≤ r i ≤ μ i ( 1 ≤ i ≤ N − 2 ) . (25) \mu_{i+2}\le r_i\le\mu_i \qquad(1\le i\le N-2). \tag*{(25)} μ i + 2 ≤ r i ≤ μ i ( 1 ≤ i ≤ N − 2 ) . (25) If 2 ≤ k ≤ N − 1 2\le k\le N-1 2 ≤ k ≤ N − 1 and μ ∈ Γ k \mu\in\Gamma_k μ ∈ Γ k , then r ∈ Γ k − 1 r\in\Gamma_{k-1} r ∈ Γ k − 1 .
Proof. For t t t away from the entries of μ \mu μ , multiply out the rational expression for R / P R/P R / P . One obtains
R ( t ) P ( t ) = ∑ i < j ( μ i − μ j ) 2 ( t − μ i ) ( t − μ j ) = ( N t − S ) ∑ i 1 t − μ i − N 2 . \frac{R(t)}{P(t)}=\sum_{i<j}\frac{(\mu_i-\mu_j)^2}{(t-\mu_i)(t-\mu_j)}=(Nt-S)\sum_i\frac{1}{t-\mu_i}-N^2. P ( t ) R ( t ) = i < j ∑ ( t − μ i ) ( t − μ j ) ( μ i − μ j ) 2 = ( N t − S ) i ∑ t − μ i 1 − N 2 .
This proves the first identity by polynomial continuation. The degree is N − 2 N-2 N − 2 and its leading coefficient is ∑ i < j ( μ i − μ j ) 2 = D 2 > 0 \sum_{i<j}(\mu_i-\mu_j)^2=D_2>0 ∑ i < j ( μ i − μ j ) 2 = D 2 > 0 .
To locate the roots, use a symmetric compression. Let D = diag ( μ ) D=\operatorname{diag}(\mu) D = diag ( μ ) and let V V V be the N N N by 2 2 2 matrix with columns ( 1 , … , 1 ) T (1,\ldots,1)^T ( 1 , … , 1 ) T and μ \mu μ . These columns are independent. Let W W W have orthonormal columns spanning their orthogonal complement. The Cauchy–Binet formula gives
P ( t ) det ( V T ( t I − D ) − 1 V ) = R ( t ) . P(t)\det\left(V^T(tI-D)^{-1}V\right)=R(t). P ( t ) det ( V T ( t I − D ) − 1 V ) = R ( t ) .
In fact, the 2 2 2 by 2 2 2 minor of V V V in rows i , j i,j i , j is μ j − μ i \mu_j-\mu_i μ j − μ i , so each term is exactly the corresponding term of R R R . Choose also an orthonormal basis Z Z Z for the range of V V V and write V = Z B V=ZB V = Z B . The complementary-minor identity for the inverse of t I − D tI-D t I − D , in the orthonormal basis ( Z , W ) (Z,W) ( Z , W ) , yields
P ( t ) det ( V T ( t I − D ) − 1 V ) = det ( V T V ) det ( t I − W T D W ) . P(t)\det\left(V^T(tI-D)^{-1}V\right)=\det(V^TV)\det(tI-W^TDW). P ( t ) det ( V T ( t I − D ) − 1 V ) = det ( V T V ) det ( t I − W T D W ) .
These identities initially hold away from finitely many values of t t t and therefore hold as polynomial identities. Since det ( V T V ) = D 2 \det(V^TV)=D_2 det ( V T V ) = D 2 , the roots r i r_i r i are the eigenvalues of the real symmetric matrix W T D W W^TDW W T D W . Restricting a symmetric quadratic form to a subspace of codimension two gives μ i + 2 ≤ r i ≤ μ i \mu_{i+2}\le r_i\le\mu_i μ i + 2 ≤ r i ≤ μ i by the min–max characterization of eigenvalues. For N = 2 N=2 N = 2 , the compressed matrix has size zero and its determinant is 1 1 1 . It remains to prove the cone assertion. Put e j = σ j ( μ ) e_j=\sigma_j(\mu) e j = σ j ( μ ) . The coefficient of ( − 1 ) j − 2 t N − j (-1)^{j-2}t^{N-j} ( − 1 ) j − 2 t N − j in R R R is
D j = ( N − j + 1 ) S e j − 1 − N j e j , 2 ≤ j ≤ N . (26) D_j=(N-j+1)Se_{j-1}-Nje_j,\qquad2\le j\le N. \tag*{(26)} D j = ( N − j + 1 ) S e j − 1 − N j e j , 2 ≤ j ≤ N . (26)
We show that D j > 0 D_j>0 D j > 0 for 2 ≤ j ≤ k + 1 2\le j\le k+1 2 ≤ j ≤ k + 1 . If e j ≤ 0 e_j\le0 e j ≤ 0 , then j = k + 1 j=k+1 j = k + 1 , while S > 0 S>0 S > 0 and e j − 1 > 0 e_{j-1}>0 e j − 1 > 0 , so the assertion follows immediately. If e j > 0 e_j>0 e j > 0 , then μ ∈ Γ j \mu\in\Gamma_j μ ∈ Γ j . Newton’s inequalities say that the positive ratios
e h / ( N h ) e h − 1 / ( N h − 1 ) ( 1 ≤ h ≤ j ) \frac{e_h/\binom{N}{h}}{e_{h-1}/\binom{N}{h-1}}\qquad(1\le h\le j) e h − 1 / ( h − 1 N ) e h / ( h N ) ( 1 ≤ h ≤ j )
are nonincreasing. The ratio for h = 2 h=2 h = 2 is strictly smaller than the ratio for h = 1 h=1 h = 1 , since
( S N ) 2 − e 2 ( N 2 ) = N ∑ i μ i 2 − S 2 N 2 ( N − 1 ) > 0. \left(\frac{S}{N}\right)^2-\frac{e_2}{\binom{N}{2}}=\frac{N\sum_i\mu_i^2-S^2}{N^2(N-1)}>0. ( N S ) 2 − ( 2 N ) e 2 = N 2 ( N − 1 ) N ∑ i μ i 2 − S 2 > 0.
Consequently
e j e j − 1 < N − j + 1 N j S , \frac{e_j}{e_{j-1}}<\frac{N-j+1}{Nj}S, e j − 1 e j < N j N − j + 1 S ,
which again gives D j > 0 D_j>0 D j > 0 . Finally σ h ( r ) = D h + 2 / D 2 > 0 \sigma_h(r)=D_{h+2}/D_2>0 σ h ( r ) = D h + 2 / D 2 > 0 for 1 ≤ h ≤ k − 1 1\le h\le k-1 1 ≤ h ≤ k − 1 .
Proposition 6.2. Let 1 ≤ k ≤ n − 1 1\le k\le n-1 1 ≤ k ≤ n − 1 and k < p < k + 1 k<p<k+1 k < p < k + 1 . Suppose H ∈ Γ ‾ k H\in\overline{\Gamma}_k H ∈ Γ k has a negative smallest eigenvalue and its k k k th largest eigenvalue is positive. Divide all eigenvalues by the latter and write the normalized vector as
H = diag ( − a , μ 1 , … , μ N ) , N = n − 1 , a > 0 , μ 1 ≥ ⋯ ≥ μ N , μ k = 1. H=\operatorname{diag}(-a,\mu_1,\ldots,\mu_N),\qquad N=n-1,\quad a>0,\quad\mu_1\ge\cdots\ge\mu_N,\quad\mu_k=1. H = diag ( − a , μ 1 , … , μ N ) , N = n − 1 , a > 0 , μ 1 ≥ ⋯ ≥ μ N , μ k = 1. Set S = ∑ i μ i S=\sum_i\mu_i S = ∑ i μ i and a 0 = ( n − p ) S / ( p N ) a_0=(n-p)S/(pN) a 0 = ( n − p ) S / ( pN ) . If a ≥ a 0 a\ge a_0 a ≥ a 0 , then Q n , k , p ( H ) ≤ 0 Q_{n,k,p}(H)\le0 Q n , k , p ( H ) ≤ 0 . Equality holds precisely when a = a 0 a=a_0 a = a 0 and every μ i = 1 \mu_i=1 μ i = 1 .
Proof. Write e j = σ j ( μ ) e_j=\sigma_j(\mu) e j = σ j ( μ ) . Since σ j ( H ) = e j − a e j − 1 \sigma_j(H)=e_j-ae_{j-1} σ j ( H ) = e j − a e j − 1 , the recurrence
e 0 = 1 , e j = σ j ( H ) + a e j − 1 e_0=1,\qquad e_j=\sigma_j(H)+ae_{j-1} e 0 = 1 , e j = σ j ( H ) + a e j − 1
shows that e j > 0 e_j>0 e j > 0 for 1 ≤ j ≤ k 1\le j\le k 1 ≤ j ≤ k . Thus μ ∈ Γ k \mu\in\Gamma_k μ ∈ Γ k and S > 0 S>0 S > 0 . The definition of Q Q Q gives
Q n , k , p ( H ) = ( − 1 ) k [ V p ( μ ) + a V p − 1 ( μ ) ] . (27) Q_{n,k,p}(H)=(-1)^k\left[V_p(\mu)+aV_{p-1}(\mu)\right]. \tag*{(27)} Q n , k , p ( H ) = ( − 1 ) k [ V p ( μ ) + a V p − 1 ( μ ) ] . (27)
Apply Lemma 4.3 with l = k l=k l = k , q = p − 1 q=p-1 q = p − 1 , and c = 1 c=1 c = 1 . It gives ( − 1 ) k V p − 1 ( μ ) < 0 (-1)^kV_{p-1}(\mu)<0 ( − 1 ) k V p − 1 ( μ ) < 0 . The right side of (27) , with μ \mu μ fixed, is therefore strictly decreasing as a a a increases. It suffices to evaluate it at a 0 a_0 a 0 .
We claim the exact identity
V p ( μ ) + a 0 V p − 1 ( μ ) = − 1 p N ∑ i < j ( μ i − μ j ) 2 V p − 2 ( μ ∖ { μ i , μ j } ) . (28) V_p(\mu)+a_0V_{p-1}(\mu)=-\frac{1}{pN}\sum_{i<j}(\mu_i-\mu_j)^2V_{p-2}(\mu\setminus\{\mu_i,\mu_j\}). \tag*{(28)} V p ( μ ) + a 0 V p − 1 ( μ ) = − pN 1 i < j ∑ ( μ i − μ j ) 2 V p − 2 ( μ ∖ { μ i , μ j }) . (28)
Here the i i i th and j j j th entries are omitted. To check the identity, expand the right-hand weighted sum and compare coefficients in (26) . It equals
∑ j = 2 N ( − 1 ) j p − j [ ( N − j + 1 ) S e j − 1 − N j e j ] . \sum_{j=2}^{N}\frac{(-1)^j}{p-j}\left[(N-j+1)Se_{j-1}-Nje_j\right]. j = 2 ∑ N p − j ( − 1 ) j [ ( N − j + 1 ) S e j − 1 − N j e j ] .
One may add the terms j = 0 , 1 , N + 1 j=0,1,N+1 j = 0 , 1 , N + 1 by setting e j = 0 e_j=0 e j = 0 outside 0 ≤ j ≤ N 0\le j\le N 0 ≤ j ≤ N ; these added coefficients vanish. Use N − j + 1 = ( p − j ) + ( n − p ) N-j+1=(p-j)+(n-p) N − j + 1 = ( p − j ) + ( n − p ) and j = p − ( p − j ) j=p-(p-j) j = p − ( p − j ) . The terms without a denominator cancel because
∑ j = 0 N ( − 1 ) j e j = ∏ i ( 1 − μ i ) = 0 , \sum_{j=0}^{N}(-1)^je_j=\prod_i(1-\mu_i)=0, j = 0 ∑ N ( − 1 ) j e j = i ∏ ( 1 − μ i ) = 0 ,
and the remaining terms are − p N V p ( μ ) − ( n − p ) S V p − 1 ( μ ) -pNV_p(\mu)-(n-p)SV_{p-1}(\mu) − pN V p ( μ ) − ( n − p ) S V p − 1 ( μ ) . This proves (28) .
If μ \mu μ is constant, all its entries equal 1 1 1 and the weighted sum vanishes. Suppose it is not constant and write R ( t ) = D 2 ∏ ( t − r i ) R(t)=D_2\prod(t-r_i) R ( t ) = D 2 ∏ ( t − r i ) as in Lemma 6.1 . Comparing the coefficients once more shows that the sum in (28) is
D 2 V p − 2 ( r ) . D_2V_{p-2}(r). D 2 V p − 2 ( r ) .
For 2 ≤ k ≤ n − 2 2\le k\le n-2 2 ≤ k ≤ n − 2 , the root vector belongs to Γ k − 1 \Gamma_{k-1} Γ k − 1 . Moreover, (25) and μ k = 1 \mu_k=1 μ k = 1 imply
r k ≤ 1 ≤ r k − 2 , r_k\le1\le r_{k-2}, r k ≤ 1 ≤ r k − 2 ,
with nonexistent endpoints omitted. These are exactly the restrictions (22) with l = k − 1 l=k-1 l = k − 1 and b = 1 b=1 b = 1 . The ( k − 1 ) (k-1) ( k − 1 ) st largest root is positive by the open-cone conclusion of Lemma 2.1 . Lemma 4.3 , with q = p − 2 q=p-2 q = p − 2 , gives ( − 1 ) k V p − 2 ( r ) > 0 (-1)^kV_{p-2}(r)>0 ( − 1 ) k V p − 2 ( r ) > 0 . Combining this sign with (27) –(28) proves strict negativity at a 0 a_0 a 0 .
The two end values of k k k can be handled directly. If k = 1 k=1 k = 1 , then p − 2 ∈ ( − 1 , 0 ) p-2\in(-1,0) p − 2 ∈ ( − 1 , 0 ) and every r i ≤ μ 1 = 1 r_i\le\mu_1=1 r i ≤ μ 1 = 1 . Termwise integration gives
V p − 2 ( r ) = − ∫ 0 1 t 1 − p ∏ i ( 1 − t r i ) d t < 0. V_{p-2}(r)=-\int_0^1t^{1-p}\prod_i(1-tr_i)\,\mathrm{d}t<0. V p − 2 ( r ) = − ∫ 0 1 t 1 − p i ∏ ( 1 − t r i ) d t < 0.
The formula also holds for an empty root vector. If k = n − 1 = N k=n-1=N k = n − 1 = N , then every μ i ≥ 1 \mu_i\ge1 μ i ≥ 1 . Any vector ν \nu ν obtained by deleting two entries has length n − 3 n-3 n − 3 and satisfies
V p − 2 ( ν ) = ∫ 0 1 t p − n ∏ i = 1 n − 3 ( t − ν i ) d t . V_{p-2}(\nu)=\int_0^1t^{p-n}\prod_{i=1}^{n-3}(t-\nu_i)\,\mathrm{d}t. V p − 2 ( ν ) = ∫ 0 1 t p − n i = 1 ∏ n − 3 ( t − ν i ) d t .
This integral has strict sign ( − 1 ) n − 3 = ( − 1 ) k (-1)^{n-3}=(-1)^k ( − 1 ) n − 3 = ( − 1 ) k ; it converges because p > n − 1 p>n-1 p > n − 1 . At least one pair has positive weight when μ \mu μ is nonconstant. The weighted sum therefore has the same strict sign.
We have proved strict negativity at a 0 a_0 a 0 in every nonconstant case, and equality there in the constant case. Strict decrease in a a a proves the asserted inequality and equality characterization.
The following elementary observation turns a strict matrix sign into an estimate with uniform constants. We will use it for both coefficient classes.
Lemma 6.3. Let C C C be a closed cone in a finite-dimensional normed space. Let Q Q Q and F F F be continuous on C C C , positively homogeneous of degrees p > 0 p>0 p > 0 and one, respectively. Suppose
H ∈ C ∖ { 0 } , F ( H ) ≤ 0 ⟹ Q ( H ) < 0. H\in C\setminus\{0\},\qquad F(H)\le0\quad\Longrightarrow\quad Q(H)<0. H ∈ C ∖ { 0 } , F ( H ) ≤ 0 ⟹ Q ( H ) < 0. There are C , c > 0 C,c>0 C , c > 0 such that
Q ( H ) ≤ C ( F ( H ) + ) p − c ∣ H ∣ p ( H ∈ C ) . Q(H)\le C(F(H)^+)^p-c|H|^p\qquad(H\in C). Q ( H ) ≤ C ( F ( H ) + ) p − c ∣ H ∣ p ( H ∈ C ) . Proof. On the compact set K = C ∩ { ∣ H ∣ = 1 } K=C\cap\{|H|=1\} K = C ∩ { ∣ H ∣ = 1 } , first suppose { F ≤ 0 } \{F\le0\} { F ≤ 0 } is nonempty. Its strict negative maximum for Q Q Q , and continuity, give δ , ε > 0 \delta,\varepsilon>0 δ , ε > 0 such that F ≤ ε F\le\varepsilon F ≤ ε implies Q ≤ − δ Q\le-\delta Q ≤ − δ on K K K . Choose 0 < c < δ 0<c<\delta 0 < c < δ . On the remaining part of K K K , F > ε F>\varepsilon F > ε and Q Q Q is bounded above, so a sufficiently large C C C gives Q + c ≤ C ( F + ) p Q+c\le C(F^+)^p Q + c ≤ C ( F + ) p . If { F ≤ 0 } \{F\le0\} { F ≤ 0 } is empty, F F F has a positive minimum on K K K and the same estimate follows directly. Homogeneity proves the assertion.
Proposition 6.4. Let 0 < δ ≤ 1 / n 0<\delta\le1/n 0 < δ ≤ 1/ n , k < p < k + 1 k<p<k+1 k < p < k + 1 , and p > n [ 1 − ( n − 1 ) δ ] p>n[1-(n-1)\delta] p > n [ 1 − ( n − 1 ) δ ] . There are constants C , c > 0 C,c>0 C , c > 0 , depending only on n , k , p , δ n,k,p,\delta n , k , p , δ , such that
Q n , k , p ( H ) ≤ C ( F δ ( H ) + ) p − c ∣ H ∣ p ( H ∈ Γ ‾ k ) . (29) Q_{n,k,p}(H)\le C\bigl(F_\delta(H)^+\bigr)^p-c|H|^p\qquad(H\in\overline{\Gamma}_k). \tag*{(29)} Q n , k , p ( H ) ≤ C ( F δ ( H ) + ) p − c ∣ H ∣ p ( H ∈ Γ k ) . (29) In particular the last term may be replaced by − c ′ σ k ( H ) p / k -c'\sigma_k(H)^{p/k} − c ′ σ k ( H ) p / k for a suitable c ′ > 0 c'>0 c ′ > 0 .
Proof. First suppose H ≠ 0 H \ne0 H = 0 and F δ ( H ) ≤ 0 F_\delta(H) \le0 F δ ( H ) ≤ 0 . A nonzero positive semidefinite matrix has F δ ( H ) > 0 F_\delta(H) > 0 F δ ( H ) > 0 . Thus H H H has a negative smallest eigenvalue, and Lemma 2.1 implies that the selected eigenvalue b b b is positive. Normalize b = 1 b = 1 b = 1 and use the notation of Proposition 6.2. The condition F δ ( H ) ≤ 0 F_\delta(H) \le0 F δ ( H ) ≤ 0 becomes
0 ≥ F δ ( H ) = δ S − [ 1 − ( n − 1 ) δ ] a . 0 \ge F_\delta(H) = \delta S - [1 - (n - 1)\delta]a. 0 ≥ F δ ( H ) = δ S − [ 1 − ( n − 1 ) δ ] a .
Since S > 0 S > 0 S > 0 ,
a ≥ δ S 1 − ( n − 1 ) δ > ( n − p ) S p ( n − 1 ) = a 0 . a \ge\frac{\delta S}{1 - (n - 1)\delta} > \frac{(n-p)S}{p(n-1)} = a_0. a ≥ 1 − ( n − 1 ) δ δ S > p ( n − 1 ) ( n − p ) S = a 0 .
The strict inequality is equivalent to the assumed restriction on p p p . Strict decrease in (27) therefore gives Q n , k , p ( H ) < 0 Q_{n,k,p}(H) < 0 Q n , k , p ( H ) < 0 .
Lemma 6.3 , applied on Γ ‾ k \overline{\Gamma}_k Γ k , now gives (29) . Finally, σ k ( H ) p / k ≤ C n , k , p ∣ H ∣ p \sigma_k(H)^{p/k} \le C_{n,k,p}|H|^p σ k ( H ) p / k ≤ C n , k , p ∣ H ∣ p .
The matrix inequality for general ellipticity bounds
Fix 1 ≤ k < n 1 \le k < n 1 ≤ k < n and k < p < k + 1 k < p < k + 1 k < p < k + 1 . We prove the matrix sign for M 1 , ρ − \mathcal{M}^{-}_{1,\rho} M 1 , ρ − by reducing the spectrum to two distinct nonzero values. The negative eigenvalues can be averaged. The positive eigenvalues need a separate argument, since moving them also changes the selected eigenvalue in the definition of Q Q Q . The construction of Q d , k , p Q_{d,k,p} Q d , k , p and the results of § 5 apply in each dimension d ≥ k + 1 d \ge k + 1 d ≥ k + 1 . Eigenvalues in this section are ordered increasingly unless stated otherwise.
Lemma 7.1. Suppose H ∈ Γ ‾ k ⊂ Sym ( d ) H \in\overline{\Gamma}_k \subset\operatorname{Sym}(d) H ∈ Γ k ⊂ Sym ( d ) has a negative eigenvalue. Increasing the magnitude of one negative eigenvalue strictly decreases Q d , k , p ( H ) Q_{d,k,p}(H) Q d , k , p ( H ) as long as the resulting matrix remains in Γ ‾ k \overline{\Gamma}_k Γ k . Replacing two negative eigenvalues by their arithmetic mean preserves Γ ‾ k \overline{\Gamma}_k Γ k , the selected eigenvalue, and the positive and negative spectral sums. This replacement strictly increases Q Q Q unless the entries are equal. Consequently, replacing all negative entries by their common mean does not decrease Q Q Q .
Proof. The selected eigenvalue is positive by Lemma 2.1 ; normalize it to one. Write the spectrum as ( − a , ξ ) (-a,\xi) ( − a , ξ ) , where a > 0 a > 0 a > 0 . Deleting the negative entry gives ξ ∈ Γ ‾ k \xi\in\overline{\Gamma}_k ξ ∈ Γ k , and its k k k th largest entry is still one. Multiaffinity of the elementary symmetric functions gives
Q d , k , p ( − a , ξ ) = ( − 1 ) k { V p ( ξ ) + a V p − 1 ( ξ ) } . Q_{d,k,p}(-a,\xi) = (-1)^k\{V_p(\xi) + aV_{p-1}(\xi)\}. Q d , k , p ( − a , ξ ) = ( − 1 ) k { V p ( ξ ) + a V p − 1 ( ξ )} .
Lemma 4.3 , applied with index k k k and exponent p − 1 p - 1 p − 1 , shows that
∂ ∂ a Q d , k , p ( − a , ξ ) = ( − 1 ) k V p − 1 ( ξ ) < 0. \frac{\partial}{\partial a}Q_{d,k,p}(-a,\xi) = (-1)^k V_{p-1}(\xi) < 0. ∂ a ∂ Q d , k , p ( − a , ξ ) = ( − 1 ) k V p − 1 ( ξ ) < 0.
For two negative entries write the spectrum as ( − a , − c , ξ ) (-a,-c,\xi) ( − a , − c , ξ ) and keep s = a + c s = a+c s = a + c fixed. Deleting both entries gives ξ ∈ Γ ‾ k \xi\in\overline{\Gamma}_k ξ ∈ Γ k , with k k k th largest entry one. We have
σ j ( H ) = σ j ( ξ ) − s σ j − 1 ( ξ ) + a c σ j − 2 ( ξ ) , \sigma_j(H) = \sigma_j(\xi) - s\sigma_{j-1}(\xi) + ac\sigma_{j-2}(\xi), σ j ( H ) = σ j ( ξ ) − s σ j − 1 ( ξ ) + a c σ j − 2 ( ξ ) ,
Q d , k , p ( H ) = ( − 1 ) k { V p ( ξ ) + s V p − 1 ( ξ ) + a c V p − 2 ( ξ ) } . Q_{d,k,p}(H) = (-1)^k\{V_p(\xi) + sV_{p-1}(\xi) + acV_{p-2}(\xi)\}. Q d , k , p ( H ) = ( − 1 ) k { V p ( ξ ) + s V p − 1 ( ξ ) + a c V p − 2 ( ξ )} .
Symmetric functions with negative indices are zero. Increasing a c ac a c preserves σ j ( H ) ≥ 0 \sigma_j(H) \ge0 σ j ( H ) ≥ 0 for j ≤ k j \le k j ≤ k . For k ≥ 2 k \ge2 k ≥ 2 , Lemma 4.3 with index k − 1 k - 1 k − 1 , exponent p − 2 p - 2 p − 2 , and cutoff one gives ( − 1 ) k V p − 2 ( ξ ) > 0 (-1)^kV_{p-2}(\xi) > 0 ( − 1 ) k V p − 2 ( ξ ) > 0 . The cutoff is permitted because it is the k k k th largest entry of ξ \xi ξ . For k = 1 k = 1 k = 1 , every entry of ξ \xi ξ is at most one and q = p − 2 ∈ ( − 1 , 0 ) q = p - 2 \in(-1,0) q = p − 2 ∈ ( − 1 , 0 ) , so
V q ( ξ ) = − ∫ 0 1 t − q − 1 ∏ i ( 1 − t ξ i ) d t < 0. V_q(\xi) = -\int_0^1 t^{-q-1}\prod_i(1-t\xi_i)\,dt < 0. V q ( ξ ) = − ∫ 0 1 t − q − 1 i ∏ ( 1 − t ξ i ) d t < 0.
This gives the same positive coefficient of a c ac a c . The arithmetic mean maximizes a c ac a c at fixed s s s , proving the assertion for two entries. Repeatedly averaging the least and greatest negative magnitudes makes their variance tend to zero. Continuity proves the assertion for all negative entries.
Lemma 7.2. Fix a nonempty list of negative entries, a list of zero entries, and the sum of the remaining positive entries. Allow the latter entries to be nonnegative and require the full spectrum to belong to Γ ‾ k \overline{\Gamma}_k Γ k . Suppose that this compact set is nonempty and that every global maximizer of Q d , k , p Q_{d,k,p} Q d , k , p belongs to Γ k \Gamma_k Γ k and has every variable entry positive. Then some global maximizer has all variable entries equal.
Proof. The negative entries are fixed, so the selected eigenvalue b b b stays positive on the feasible set. Temporarily regard b b b as an independent variable and write
F ( λ , b ) = ( − 1 ) k ∑ j = 0 d ( − 1 ) j σ j ( λ ) b p − j p − j . F(\lambda,b)=(-1)^k\sum_{j=0}^{d}\frac{(-1)^j\sigma_j(\lambda)b^{p-j}}{p-j}. F ( λ , b ) = ( − 1 ) k j = 0 ∑ d p − j ( − 1 ) j σ j ( λ ) b p − j . This function is symmetric and affine in each eigenvalue separately. Differentiation in b b b gives
F b ( λ , b ) = ( − 1 ) k b p − d − 1 ∏ h = 1 d ( b − λ h ) . F_b(\lambda,b)=(-1)^k b^{p-d-1}\prod_{h=1}^{d}(b-\lambda_h). F b ( λ , b ) = ( − 1 ) k b p − d − 1 h = 1 ∏ d ( b − λ h ) . Thus F b ( λ , b ( λ ) ) = 0 F_b(\lambda,b(\lambda))=0 F b ( λ , b ( λ )) = 0 at the selected eigenvalue. The selection b ( λ ) b(\lambda) b ( λ ) is locally Lipschitz, and F F F is smooth where b > 0 b>0 b > 0 . In a first-order expansion, the variation of the cutoff therefore contributes only O ( ∣ h ∣ 2 ) O(|h|^2) O ( ∣ h ∣ 2 ) . It follows that Q ( λ ) = F ( λ , b ( λ ) ) Q(\lambda)=F(\lambda,b(\lambda)) Q ( λ ) = F ( λ , b ( λ )) is C 1 C^1 C 1 , including at repeated selected eigenvalues, with Q i = F i ( λ , b ( λ ) ) Q_i=F_i(\lambda,b(\lambda)) Q i = F i ( λ , b ( λ )) .
Assume that no maximizer has all variable entries equal. Choose a maximizer where the multiplicity r r r of b b b is least. The cone and positivity constraints are inactive at this point, so F i = F j F_i=F_j F i = F j for every pair of variable entries. Symmetry and multiaffinity give the exact identity
F i − F j = ( λ j − λ i ) F i j . F_i-F_j=(\lambda_j-\lambda_i)F_{ij}. F i − F j = ( λ j − λ i ) F ij . For fixed b b b , the derivative F i j F_{ij} F ij does not depend on λ i \lambda_i λ i or λ j \lambda_j λ j . Choose i i i with λ i = b \lambda_i=b λ i = b and j j j with λ j ≠ b \lambda_j\ne b λ j = b . Then F i j = 0 F_{ij}=0 F ij = 0 .
First suppose r = 1 r=1 r = 1 . Under the sum-preserving variation λ i ( t ) = b + t \lambda_i(t)=b+t λ i ( t ) = b + t and λ j ( t ) = λ j − t \lambda_j(t)=\lambda_j-t λ j ( t ) = λ j − t , the selected eigenvalue is b ( t ) = b + t b(t)=b+t b ( t ) = b + t for small ∣ t ∣ |t| ∣ t ∣ . At t = 0 t=0 t = 0 ,
F b b = ( − 1 ) k b p − d − 1 ∏ h ≠ i ( b − λ h ) < 0 , F i b = − F b b , F j b = 0. F_{bb}=(-1)^k b^{p-d-1}\prod_{h\ne i}(b-\lambda_h)<0,\qquad F_{ib}=-F_{bb},\qquad F_{jb}=0. F bb = ( − 1 ) k b p − d − 1 h = i ∏ ( b − λ h ) < 0 , F ib = − F bb , F j b = 0. There are exactly k − 1 k-1 k − 1 entries greater than the simple selected eigenvalue, which proves the displayed sign. Since F i i = F j j = 0 F_{ii}=F_{jj}=0 F ii = F j j = 0 ,
d 2 d t 2 Q ( λ ( t ) ) ∣ t = 0 = − 2 F i j + 2 ( F i b − F j b ) + F b b = − F b b > 0. \left.\frac{d^2}{dt^2}Q(\lambda(t))\right|_{t=0}=-2F_{ij}+2(F_{ib}-F_{jb})+F_{bb}=-F_{bb}>0. d t 2 d 2 Q ( λ ( t )) t = 0 = − 2 F ij + 2 ( F ib − F j b ) + F bb = − F bb > 0. This contradicts maximality.
Now suppose r ≥ 2 r\ge2 r ≥ 2 . Keep the cutoff fixed at b b b in the same variation. The change of F F F is an exact quadratic polynomial, whose linear and quadratic coefficients are F i − F j F_i-F_j F i − F j and − F i j -F_{ij} − F ij . Both vanish, so F ( λ ( t ) , b ) F(\lambda(t),b) F ( λ ( t ) , b ) is constant. To choose a sign of t t t for which the selected eigenvalue remains b b b , list the entries temporarily in decreasing order. Let the block equal to b b b occupy positions L , … , H L,\ldots,H L , … , H , where L ≤ k ≤ H L\le k\le H L ≤ k ≤ H . If k > L k>L k > L , take t > 0 t>0 t > 0 ; if k = L k=L k = L , take t < 0 t<0 t < 0 . For sufficiently small nonzero t t t , the selected eigenvalue remains b b b and has multiplicity r − 1 r-1 r − 1 . The other changed entry remains separated from b b b , and the cone and positivity conditions remain satisfied. We have found a maximizer with smaller multiplicity, a contradiction.
For later use put
H d , m , a = diag ( − a , … , − a ⏟ m , 1 , … , 1 ⏟ d − m ) , a > 0. (30) H_{d,m,a}=\operatorname{diag}(\underbrace{-a,\ldots,-a}_{m},\underbrace{1,\ldots,1}_{d-m}),\qquad a>0. \tag*{(30)} H d , m , a = diag ( m − a , … , − a , d − m 1 , … , 1 ) , a > 0. (30)
Proposition 7.3. Suppose that for every k + 1 ≤ d ≤ n k+1\le d\le n k + 1 ≤ d ≤ n , 1 ≤ m ≤ d − k 1\le m\le d-k 1 ≤ m ≤ d − k , and a > 0 a>0 a > 0 satisfying
H d , m , a ∈ Γ ‾ k , ( d − m ) − ρ m a ≤ 0 , H_{d,m,a}\in\overline{\Gamma}_k,\qquad(d-m)-\rho ma\le0, H d , m , a ∈ Γ k , ( d − m ) − ρ ma ≤ 0 , one has Q d , k , p ( H d , m , a ) < 0 Q_{d,k,p}(H_{d,m,a})<0 Q d , k , p ( H d , m , a ) < 0 . Then Q n , k , p ( H ) < 0 Q_{n,k,p}(H)<0 Q n , k , p ( H ) < 0 for every nonzero H ∈ Γ ‾ k H\in\overline{\Gamma}_k H ∈ Γ k with M 1 , ρ − ( H ) ≤ 0 M_{1,\rho}^{-}(H)\le0 M 1 , ρ − ( H ) ≤ 0 .
Proof. We use induction on the dimension, starting at d = k + 1 d=k+1 d = k + 1 . A nonzero positive semidefinite matrix has M 1 , ρ − ( H ) = tr H > 0 M_{1,\rho}^{-}(H)=\operatorname{tr}H>0 M 1 , ρ − ( H ) = tr H > 0 . Thus every spectrum under consideration has a negative entry, a positive selected eigenvalue, and at least k k k positive entries. The first part of the proof of Proposition 4.2 gives Q ( H ) < 0 Q(H)<0 Q ( H ) < 0 when σ k ( H ) = 0 \sigma_k(H)=0 σ k ( H ) = 0 . By Newton’s inequalities, a point of Γ ‾ k \overline{\Gamma}_k Γ k with σ k > 0 \sigma_k>0 σ k > 0 is in Γ k \Gamma_k Γ k .
Deleting zero entries preserves cone membership, the selected eigenvalue, both spectral sums, and Q Q Q . In fact, every symmetric function of degree exceeding the new dimension vanishes, and all remaining terms in (17) are unchanged. A face with a zero entry therefore reduces to a lower dimension with the same k , p , ρ k,p,\rho k , p , ρ . The remaining dimension is at least k + 1 k+1 k + 1 . The induction hypothesis handles these faces; none occur in the base dimension.
Suppose an interior-cone spectrum without zero entries satisfies Q ≥ 0 Q\ge0 Q ≥ 0 and M 1 , ρ − ≤ 0 M_{1,\rho}^{-}\le0 M 1 , ρ − ≤ 0 . Let m m m be its number of negative entries. Average them to − a -a − a by Lemma 7.1 . This does not decrease Q Q Q and preserves both spectral sums. Let S S S be the positive sum. Maximize Q Q Q over
{ x ∈ [ 0 , S ] d − m : ∑ i x i = S , ( ( − a ) m , x ) ∈ Γ ‾ k } . \left\{x\in[0,S]^{d-m}:\ \sum_i x_i=S,\ ((-a)^m,x)\in\overline{\Gamma}_k\right\}. { x ∈ [ 0 , S ] d − m : i ∑ x i = S , (( − a ) m , x ) ∈ Γ k } .
The maximum is nonnegative, and all spectra in this set satisfy M 1 , ρ − = S − ρ m a ≤ 0 M_{1,\rho}^{-}=S-\rho ma\le0 M 1 , ρ − = S − ρ ma ≤ 0 . A maximizer cannot have σ k = 0 \sigma_k=0 σ k = 0 or a zero variable entry, by the preceding two paragraphs. Lemma 7.2 therefore gives a maximizer with all positive entries equal to t = S / ( d − m ) > 0 t=S/(d-m)>0 t = S / ( d − m ) > 0 . Homogeneity yields
Q d , k , p ( ( − a ) m , t d − m ) = t p Q d , k , p ( H d , m , a / t ) < 0 , Q_{d,k,p}((-a)^m,t^{d-m})=t^pQ_{d,k,p}(H_{d,m,a/t})<0, Q d , k , p (( − a ) m , t d − m ) = t p Q d , k , p ( H d , m , a / t ) < 0 ,
contrary to the nonnegative maximum. □
Krawtchouk polynomials and the critical exponent
The preceding reduction leaves a finite family of one-variable sign questions. We identify their critical values explicitly. For d ≥ 2 d\ge2 d ≥ 2 , 1 ≤ m < d 1\le m<d 1 ≤ m < d , and a > 0 a>0 a > 0 , define
R m ( x , a ) = m ! [ z m ] ( 1 + a z ) x ( 1 − z ) d − x . (31) R_m(x,a)=m^x(1-z)^{d-x}. \tag*{(31)} R m ( x , a ) = m ! [ z m ] ( 1 + a z ) x ( 1 − z ) d − x . (31)
Here [ z m ] [z^m] [ z m ] denotes the Taylor coefficient at zero. The result is a polynomial in x x x of degree m m m , with leading coefficient ( 1 + a ) m (1+a)^m ( 1 + a ) m . In the standard notation for Krawtchouk polynomials,
R m ( x , a ) = ( − 1 ) m m ! ( d m ) K m ( x ; 1 1 + a , d ) ; R_m(x,a)=(-1)^m m!\binom{d}{m}K_m\left(x;\frac{1}{1+a},d\right); R m ( x , a ) = ( − 1 ) m m ! ( m d ) K m ( x ; 1 + a 1 , d ) ;
see [27 Eq. 18.23.3] . The generating-function identity first holds for x = 0 , … , d x=0,\ldots,d x = 0 , … , d and then for every x x x by polynomial identity.
Let K d , m ( a ) K_{d,m}(a) K d , m ( a ) be the tridiagonal matrix on j = 0 , … , m − 1 j=0,\ldots,m-1 j = 0 , … , m − 1 with rates
λ j ( a ) = d − j 1 + a , μ j ( a ) = j a 1 + a . (32) \lambda_j(a)=\frac{d-j}{1+a},\qquad\mu_j(a)=\frac{ja}{1+a}. \tag*{(32)} λ j ( a ) = 1 + a d − j , μ j ( a ) = 1 + a j a . (32)
Its diagonal, superdiagonal, and subdiagonal are respectively λ j + μ j \lambda_j+\mu_j λ j + μ j , − λ j -\lambda_j − λ j , and − μ j -\mu_j − μ j . At a = ( d − m ) / ( m ρ ) a=(d-m)/(m\rho) a = ( d − m ) / ( m ρ ) it is K d , m , ρ K_{d,m,\rho} K d , m , ρ from (7) .
Lemma 8.1. The roots of R m ( ⋅ , a ) R_m(\cdot,a) R m ( ⋅ , a ) are simple and satisfy 0 < κ 1 < ⋯ < κ m < d 0<\kappa_1<\cdots<\kappa_m<d 0 < κ 1 < ⋯ < κ m < d . Between any two consecutive roots there is an integer strictly between them. Moreover,
( − 1 ) m R m ( j , a ) = m ! ( d m ) ( d j ) σ j ( H d , m , a ) , 0 ≤ j ≤ d , (33) (-1)^m R_m(j,a)=\frac{m!\binom{d}{m}}{\binom{d}{j}}\sigma_j(H_{d,m,a}),\qquad0\le j\le d, \tag*{(33)} ( − 1 ) m R m ( j , a ) = ( j d ) m ! ( m d ) σ j ( H d , m , a ) , 0 ≤ j ≤ d , (33) and
det ( K d , m ( a ) − x I ) = ( − 1 ) m R m ( x , a ) ( 1 + a ) m . (34) \det(K_{d,m}(a)-xI)=\frac{(-1)^mR_m(x,a)}{(1+a)^m}. \tag*{(34)} det ( K d , m ( a ) − x I ) = ( 1 + a ) m ( − 1 ) m R m ( x , a ) . (34) The smallest eigenvalue of K d , m ( a ) K_{d,m}(a) K d , m ( a ) is therefore positive and simple and has a strictly positive right eigenvector.
Proof. The Krawtchouk polynomials are orthogonal on { 0 , … , d } \{0,\ldots,d\} { 0 , … , d } with positive weights
w j = ( d j ) a d − j ( 1 + a ) d . w_j=\binom{d}{j}\frac{a^{d-j}}{(1+a)^d}. w j = ( j d ) ( 1 + a ) d a d − j .
The zeros of a degree-m m m orthogonal polynomial are simple and lie in the open convex hull of the support. See [27 Sections 18.19 and 18.2(vi)] . Separation by support points follows from orthogonality: if two consecutive zeros had no support point strictly between them, division by the two corresponding linear factors would give a polynomial of degree m − 2 m-2 m − 2 whose product with R m R_m R m is nonnegative at every support point and positive at some such point, contradicting orthogonality. The same argument allows zeros at support points.
Changing z z z to − z -z − z in (31) gives
( − 1 ) m R m ( j , a ) m ! = ∑ q ( − a ) q ( j q ) ( d − j m − q ) . \frac{(-1)^mR_m(j,a)}{m!}=\sum_q(-a)^q\binom{j}{q}\binom{d-j}{m-q}. m ! ( − 1 ) m R m ( j , a ) = q ∑ ( − a ) q ( q j ) ( m − q d − j ) .
Use the combinatorial identity
( d j ) ( j q ) ( d − j m − q ) = ( d m ) ( m q ) ( d − m j − q ) . \binom{d}{j}\binom{j}{q}\binom{d-j}{m-q}=\binom{d}{m}\binom{m}{q}\binom{d-m}{j-q}. ( j d ) ( q j ) ( m − q d − j ) = ( m d ) ( q m ) ( j − q d − m ) .
The sum on the right after this substitution is ( d m ) σ j ( H d , m , a ) \binom{d}{m}\sigma_j(H_{d,m,a}) ( m d ) σ j ( H d , m , a ) , proving (33) .
For the determinant identity put B = ( 1 + a ) x − d B=(1+a)x-d B = ( 1 + a ) x − d and F ( z ) = ( 1 + a z ) x ( 1 − z ) d − x F(z)=(1+az)^x(1-z)^{d-x} F ( z ) = ( 1 + a z ) x ( 1 − z ) d − x . Logarithmic differentiation gives
[ 1 + ( a − 1 ) z − a z 2 ] F ′ ( z ) = [ B − a d z ] F ( z ) . [1+(a-1)z-az^2]F'(z)=[B-adz]F(z). [ 1 + ( a − 1 ) z − a z 2 ] F ′ ( z ) = [ B − a d z ] F ( z ) .
Comparison of coefficients yields R 0 = 1 R_0=1 R 0 = 1 , R 1 = B R_1=B R 1 = B , and
R j = [ B − ( a − 1 ) ( j − 1 ) ] R j − 1 − a ( j − 1 ) ( d − j + 2 ) R j − 2 ( j ≥ 2 ) . R_j=[B-(a-1)(j-1)]R_{j-1}-a(j-1)(d-j+2)R_{j-2}\qquad(j\ge2). R j = [ B − ( a − 1 ) ( j − 1 )] R j − 1 − a ( j − 1 ) ( d − j + 2 ) R j − 2 ( j ≥ 2 ) .
The determinants of the leading blocks of ( 1 + a ) ( K d , m ( a ) − x I ) (1+a)(K_{d,m}(a)-xI) ( 1 + a ) ( K d , m ( a ) − x I ) satisfy this recurrence after multiplication by ( − 1 ) j (-1)^j ( − 1 ) j , with initial values 1 1 1 and − B -B − B . This proves (34) .
Finally let π j = ( d j ) a − j \pi_j=\binom{d}{j}a^{-j} π j = ( j d ) a − j . The identity π j λ j = π j + 1 μ j + 1 \pi_j\lambda_j=\pi_{j+1}\mu_{j+1} π j λ j = π j + 1 μ j + 1 shows that positive diagonal similarity makes K d , m ( a ) K_{d,m}(a) K d , m ( a ) symmetric with strictly negative adjacent entries. Subtract this symmetric matrix from a sufficiently large scalar multiple of the identity. The resulting nonnegative irreducible matrix has a strictly positive eigenvector for its largest eigenvalue. This is the eigenvector for the smallest eigenvalue of K d , m ( a ) K_{d,m}(a) K d , m ( a ) , after reversing the similarity. □
Proposition 8.2. For 1 ≤ k ≤ d − 1 1\le k\le d-1 1 ≤ k ≤ d − 1 ,
H d , m , a ∈ Γ ‾ k ⟺ κ 1 ≥ k . H_{d,m,a}\in\overline{\Gamma}_k\quad\Longleftrightarrow\quad\kappa_1\ge k. H d , m , a ∈ Γ k ⟺ κ 1 ≥ k . If also k < p < k + 1 k<p<k+1 k < p < k + 1 and k ≤ d − m k\le d-m k ≤ d − m , then under this cone condition
sgn Q d , k , p ( H d , m , a ) = sgn ( κ 1 − p ) . (35) \operatorname{sgn}Q_{d,k,p}(H_{d,m,a})=\operatorname{sgn}(\kappa_1-p). \tag*{(35)} sgn Q d , k , p ( H d , m , a ) = sgn ( κ 1 − p ) . (35)
Proof. By (33) , the cone condition is equivalent to ( − 1 ) m R m ( j , a ) ≥ 0 (-1)^m R_m(j,a) \ge0 ( − 1 ) m R m ( j , a ) ≥ 0 for 1 ≤ j ≤ k 1 \le j \le k 1 ≤ j ≤ k . If κ 1 ≥ k \kappa_1 \ge k κ 1 ≥ k , all factors j − κ i j-\kappa_i j − κ i are nonpositive, so these inequalities hold. Conversely suppose κ 1 < k \kappa_1 < k κ 1 < k . If m = 1 m=1 m = 1 , or if m ≥ 2 m \ge2 m ≥ 2 and k < κ 2 k < \kappa_2 k < κ 2 , the sign at j = k j=k j = k is wrong. Otherwise κ 2 ≤ k \kappa_2 \le k κ 2 ≤ k , and Lemma 8.1 gives an integer strictly between κ 1 \kappa_1 κ 1 and κ 2 \kappa_2 κ 2 . That integer is at most k k k and again gives the wrong sign.
Under the cone condition κ 1 ≥ k \kappa_1 \ge k κ 1 ≥ k . If m ≥ 2 m \ge2 m ≥ 2 , separation by an integer implies κ 2 > k + 1 \kappa_2 > k+1 κ 2 > k + 1 . Indeed, if κ 1 < k + 1 \kappa_1 < k+1 κ 1 < k + 1 , the separating integer is at least k + 1 k+1 k + 1 ; if κ 1 ≥ k + 1 \kappa_1 \ge k+1 κ 1 ≥ k + 1 , the conclusion is immediate. Thus only the first zero can lie in ( k , k + 1 ) (k,k+1) ( k , k + 1 ) .
Write D d ( p ) = ∏ j = 0 d ( p − j ) D_d(p)=\prod_{j=0}^{d}(p-j) D d ( p ) = ∏ j = 0 d ( p − j ) and ℓ = d − m \ell=d-m ℓ = d − m . The rational functions
∑ j = 0 d ( − 1 ) j σ j ( H d , m , a ) p − j , ℓ ! ( − 1 ) ℓ R m ( p , a ) D d ( p ) \sum_{j=0}^{d}\frac{(-1)^j\sigma_j(H_{d,m,a})}{p-j}, \qquad\frac{\ell!(-1)^\ell R_m(p,a)}{D_d(p)} j = 0 ∑ d p − j ( − 1 ) j σ j ( H d , m , a ) , D d ( p ) ℓ ! ( − 1 ) ℓ R m ( p , a )
have the same residues at every pole. This follows from (33) and D d ′ ( j ) = ( − 1 ) d − j j ! ( d − j ) ! D_d'(j)=(-1)^{d-j}j!(d-j)! D d ′ ( j ) = ( − 1 ) d − j j ! ( d − j )! . Their difference has no poles and vanishes at infinity, so it is zero. The selected eigenvalue is one, and hence
Q d , k , p ( H d , m , a ) = ( − 1 ) k + d − m ( d − m ) ! R m ( p , a ) D d ( p ) . Q_{d,k,p}(H_{d,m,a})=\frac{(-1)^{k+d-m}(d-m)!R_m(p,a)}{D_d(p)}. Q d , k , p ( H d , m , a ) = D d ( p ) ( − 1 ) k + d − m ( d − m )! R m ( p , a ) .
On ( k , k + 1 ) (k,k+1) ( k , k + 1 ) , the sign of D d ( p ) D_d(p) D d ( p ) is ( − 1 ) d − k (-1)^{d-k} ( − 1 ) d − k . Consequently the sign of Q Q Q is that of ( − 1 ) m R m ( p , a ) (-1)^mR_m(p,a) ( − 1 ) m R m ( p , a ) . This is positive before the first zero, vanishes there, and is negative between the first two zeros. For m = 1 m=1 m = 1 the last interval extends to infinity. Formula (35) follows. □
Proposition 8.3. Suppose P n ( ρ ) < p < n P_n(\rho)<p<n P n ( ρ ) < p < n , p p p is not an integer, and k = ⌊ p ⌋ k=\lfloor p\rfloor k = ⌊ p ⌋ . There are C , c > 0 C,c>0 C , c > 0 , depending only on n , p , ρ n,p,\rho n , p , ρ , such that
Q n , k , p ( H ) ≤ C ( M 1 , ρ − ( H ) + ) p − c ∣ H ∣ p ( H ∈ Γ ‾ k ) . (36) Q_{n,k,p}(H)\le C\bigl(\mathcal{M}_{1,\rho}^{-}(H)^{+}\bigr)^p-c|H|^p \qquad(H\in\overline{\Gamma}_k). \tag*{(36)} Q n , k , p ( H ) ≤ C ( M 1 , ρ − ( H ) + ) p − c ∣ H ∣ p ( H ∈ Γ k ) . (36)
Proof. Consider a spectrum H d , m , a H_{d,m,a} H d , m , a from (30) in Proposition 7.3. Its extremal pairing is nonpositive precisely when a ≥ a 0 = ( d − m ) / ( m ρ ) a\ge a_0=(d-m)/(m\rho) a ≥ a 0 = ( d − m ) / ( m ρ ) . Reducing the negative magnitudes to a 0 a_0 a 0 preserves cone membership. The smallest root at a 0 a_0 a 0 is κ d , m ( ρ ) ≤ P n ( ρ ) < p \kappa_{d,m}(\rho)\le P_n(\rho)<p κ d , m ( ρ ) ≤ P n ( ρ ) < p , so Proposition 8.2 gives Q d , k , p ( H d , m , a 0 ) < 0 Q_{d,k,p}(H_{d,m,a_0})<0 Q d , k , p ( H d , m , a 0 ) < 0 . Increasing the magnitudes back to a a a stays in the cone: each intermediate spectrum is coordinatewise greater than or equal to the final admissible spectrum. Lemma 7.1 therefore gives Q d , k , p ( H d , m , a ) < 0 Q_{d,k,p}(H_{d,m,a})<0 Q d , k , p ( H d , m , a ) < 0 . Proposition 7.3 now proves the strict sign for every nonzero H ∈ Γ ‾ k H\in\overline{\Gamma}_k H ∈ Γ k with M 1 , ρ − ( H ) ≤ 0 \mathcal{M}_{1,\rho}^{-}(H)\le0 M 1 , ρ − ( H ) ≤ 0 . Apply Lemma 6.3 to obtain (36) . □
Proof of the upper estimates
We now have the three inequalities needed by Lemma 5.1: the integral sign, the bound on the whole cone, and the coercive bound for the chosen coefficient class. For Theorem 1.1, choose a noninteger q q q with P t r ( n , δ ) < q < min { p , n } P_{\mathrm{tr}}(n,\delta)<q<\min\{p,n\} P tr ( n , δ ) < q < min { p , n } , k = ⌊ q ⌋ k=\lfloor q\rfloor k = ⌊ q ⌋ .
This is possible because the threshold P t r ( n , δ ) P_{\mathrm{tr}}(n,\delta) P tr ( n , δ ) is less than n n n . Propositions 3.2, 4.2, and 6.4, with Q = Q n , k , q Q=Q_{n,k,q} Q = Q n , k , q and F = F δ F=F_\delta F = F δ , give the hypotheses of Lemma 5.1. Thus the ball estimate holds at exponent q q q .
For Theorem 1.2, Lemma 8.1 gives 0 < κ d , m ( ρ ) < d ≤ n 0<\kappa_{d,m}(\rho)<d\le n 0 < κ d , m ( ρ ) < d ≤ n , and the maximum is finite. Hence P n ( ρ ) < n P_n(\rho)<n P n ( ρ ) < n . Choose a noninteger
P n ( ρ ) < q < min { p , n } , k = ⌊ q ⌋ . P_n(\rho)<q<\min\{p,n\}, \qquad k=\lfloor q\rfloor. P n ( ρ ) < q < min { p , n } , k = ⌊ q ⌋ .
Apply the same integral and whole-cone bounds, now using Proposition 8.3 and F = M 1 , ρ − F=\mathcal{M}_{1,\rho}^{-} F = M 1 , ρ − . This proves the ball estimate for I ≤ A ≤ ρ I I\le A\le\rho I I ≤ A ≤ ρ I . For λ I ≤ A ≤ λ I \lambda I\le A\le\lambda I λ I ≤ A ≤ λ I , apply it to A / λ A/\lambda A / λ ; the forcing is divided by λ \lambda λ as well.
In both cases Hölder’s inequality gives
R 2 − n / q ∥ f ∥ L q ( B R ) ≤ ∣ B 1 ∣ 1 / q − 1 / p R 2 − n / p ∥ f ∥ L p ( B R ) . R^{2-n/q}\|f\|_{L^q(B_R)} \le|B_1|^{1/q-1/p}R^{2-n/p}\|f\|_{L^p(B_R)}. R 2 − n / q ∥ f ∥ L q ( B R ) ≤ ∣ B 1 ∣ 1/ q − 1/ p R 2 − n / p ∥ f ∥ L p ( B R ) .
For u ∈ W 2 , p ( B R ) u\in W^{2,p}(B_R) u ∈ W 2 , p ( B R ) , the inclusion in W 2 , q ( B R ) W^{2,q}(B_R) W 2 , q ( B R ) therefore proves the ball estimate at the requested exponent p p p . This includes integers and p ≥ n p\ge n p ≥ n . The intermediate q q q is chosen using only the stated parameters, so the constants have the required dependence.
It remains to pass from a ball to an arbitrary bounded open set while assuming only local Sobolev regularity. Let D = diam Ω D=\operatorname{diam}\Omega D = diam Ω , choose x 0 ∈ Ω x_0\in\Omega x 0 ∈ Ω , and set
M = sup ∂ Ω u + , v = u − M − ε , ε > 0. M=\sup_{\partial\Omega}u^+,\qquad v=u-M-\varepsilon,\qquad\varepsilon>0. M = ∂ Ω sup u + , v = u − M − ε , ε > 0.
Then Ω ⊂ B D ( x 0 ) \Omega\subset B_D(x_0) Ω ⊂ B D ( x 0 ) . Since v ≤ − ε v\le-\varepsilon v ≤ − ε on the boundary and u u u is continuous on Ω ‾ \overline{\Omega} Ω , the set { v ≥ 0 } \{v\ge0\} { v ≥ 0 } is compactly contained in Ω \Omega Ω . Choose smooth convex functions ϕ δ : R → [ 0 , ∞ ) \phi_\delta:\mathbb{R}\to[0,\infty) ϕ δ : R → [ 0 , ∞ ) with
ϕ δ = 0 on ( − ∞ , 0 ] , 0 ≤ ϕ δ ′ ≤ 1 , ϕ δ ′ ′ ≥ 0 , ϕ δ ( t ) ⟶ t + \phi_\delta=0\text{ on }(-\infty,0],\qquad0\le\phi'_\delta\le1,\qquad\phi''_\delta\ge0,\qquad\phi_\delta(t)\longrightarrow t^+ ϕ δ = 0 on ( − ∞ , 0 ] , 0 ≤ ϕ δ ′ ≤ 1 , ϕ δ ′′ ≥ 0 , ϕ δ ( t ) ⟶ t +
uniformly on bounded intervals. For example, integrate a smooth nondecreasing function which is zero on ( − ∞ , 0 ] (-\infty,0] ( − ∞ , 0 ] and one on [ δ , ∞ ) [\delta,\infty) [ δ , ∞ ) . The chain rule gives
D 2 ϕ δ ( v ) = ϕ δ ′ ( v ) D 2 u + ϕ δ ′ ′ ( v ) D u ⊗ D u . D^2\phi_\delta(v)=\phi'_\delta(v)D^2u+\phi''_\delta(v)Du\otimes Du. D 2 ϕ δ ( v ) = ϕ δ ′ ( v ) D 2 u + ϕ δ ′′ ( v ) D u ⊗ D u .
Both terms are locally in L p L^p L p . Indeed, D u ∈ L l o c 2 p Du\in L^{2p}_{\mathrm{loc}} D u ∈ L loc 2 p : if p < n p<n p < n , Sobolev embedding gives the larger exponent n p / ( n − p ) > 2 p np/(n-p)>2p n p / ( n − p ) > 2 p because p > n / 2 p>n/2 p > n /2 ; if p = n p=n p = n , every finite target exponent is available; and if p > n p>n p > n , the gradient is locally bounded. Thus the zero extension of w = ϕ δ ( v ) w=\phi_\delta(v) w = ϕ δ ( v ) belongs to W 2 , p ( B D ( x 0 ) ) W^{2,p}(B_D(x_0)) W 2 , p ( B D ( x 0 )) and has compact support in Ω \Omega Ω .
Extend A A A outside Ω \Omega Ω by I / n I/n I / n in the trace-one case and by λ I \lambda I λ I in the other case. Positivity of A A A and convexity give
( A : D 2 w ) − ≤ ϕ δ ′ ( v ) ( A : D 2 u ) − ≤ ( A : D 2 u ) − in Ω , (A:D^2w)^-\le\phi'_\delta(v)(A:D^2u)^-\le(A:D^2u)^-\quad\text{in }\Omega, ( A : D 2 w ) − ≤ ϕ δ ′ ( v ) ( A : D 2 u ) − ≤ ( A : D 2 u ) − in Ω ,
and the left side vanishes outside Ω \Omega Ω . The ball estimate for w w w , followed by δ ↓ 0 \delta\downarrow0 δ ↓ 0 and then ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , proves (2) . This completes the upper estimates of Theorems Theorem 1.1 and Theorem 1.2 .
Corollary 9.1. Assume the coefficient and exponent hypotheses of Theorem 1.1 or Theorem 1.2 . Let Ω \Omega Ω be an open set of finite volume, and let u u u belong to the class (3) . If u u u is bounded above, then
sup Ω u ≤ sup ∂ Ω u + + C ∣ Ω ∣ 2 / n − 1 / p ∥ ( A : D 2 u ) − ∥ L p ( Ω ) . (37) \sup_{\Omega}u\le\sup_{\partial\Omega}u^+ + C|\Omega|^{2/n-1/p}\|(A:D^2u)^-\|_{L^p(\Omega)}. \tag*{(37)} Ω sup u ≤ ∂ Ω sup u + + C ∣Ω ∣ 2/ n − 1/ p ∥ ( A : D 2 u ) − ∥ L p ( Ω ) . (37)
Proof. Suppose first that p < ∞ p<\infty p < ∞ . Put M = sup ∂ Ω u + M=\sup_{\partial\Omega}u^+ M = sup ∂ Ω u + and f = ( A : D 2 u ) − f=(A:D^2u)^- f = ( A : D 2 u ) − . For ε > 0 \varepsilon>0 ε > 0 , use the convex functions ϕ δ \phi_\delta ϕ δ from the preceding argument and extend w = ϕ δ ( u − M − ε ) w=\phi_\delta(u-M-\varepsilon) w = ϕ δ ( u − M − ε ) by zero. Then w ∈ W l o c 2 , p ( R n ) w\in W^{2,p}_{\mathrm{loc}}(\mathbb{R}^n) w ∈ W loc 2 , p ( R n ) , w ≥ 0 w\ge0 w ≥ 0 , and ( A : D 2 w ) − ≤ f (A:D^2w)^-\le f ( A : D 2 w ) − ≤ f on Ω \Omega Ω . The zero extension is locally valid because u − M − ε u-M-\varepsilon u − M − ε is negative near every boundary point. Extend A A A outside Ω \Omega Ω by a fixed matrix in the same coefficient class, and extend f f f by zero.
Write H = sup w H=\sup w H = sup w . If H > 0 H>0 H > 0 , choose x 0 x_0 x 0 with w ( x 0 ) > H / 2 w(x_0)>H/2 w ( x 0 ) > H /2 . Take 0 ≤ θ ≤ 1 0\le\theta\le1 0 ≤ θ ≤ 1 in C ∞ ( B R ( x 0 ) ) C^\infty(B_R(x_0)) C ∞ ( B R ( x 0 )) , with θ ( x 0 ) = 1 \theta(x_0)=1 θ ( x 0 ) = 1 , ∣ D θ ∣ ≤ C / R |D\theta|\le C/R ∣ D θ ∣ ≤ C / R , and ∣ D 2 θ ∣ ≤ C / R 2 |D^2\theta|\le C/R^2 ∣ D 2 θ ∣ ≤ C / R 2 , and put v = θ 2 w 2 v=\theta^2w^2 v = θ 2 w 2 . Since p > n / 2 p>n/2 p > n /2 , the Sobolev chain rule gives v ∈ W 2 , p ( B R ( x 0 ) ) v\in W^{2,p}(B_R(x_0)) v ∈ W 2 , p ( B R ( x 0 )) with zero boundary values. Expanding and completing the square gives
A : D 2 v = 2 θ 2 w A : D 2 w + 2 ( θ D w + 2 w D θ ) T A ( θ D w + 2 w D θ ) A:D^2v=2\theta^2wA:D^2w+2(\theta Dw+2wD\theta)^TA(\theta Dw+2wD\theta) A : D 2 v = 2 θ 2 w A : D 2 w + 2 ( θ D w + 2 w D θ ) T A ( θ D w + 2 w D θ )
+ 2 θ w 2 A : D 2 θ − 6 w 2 ( D θ ) T A D θ ≥ − 2 H f − C R − 2 H 2 1 Ω . \begin{aligned}
\qquad+2\theta w^2A:D^2\theta-6w^2(D\theta)^TAD\theta\\
&\ge-2Hf-CR^{-2}H^2\mathbf{1}_{\Omega}.
\end{aligned} + 2 θ w 2 A : D 2 θ − 6 w 2 ( D θ ) T A D θ ≥ − 2 H f − C R − 2 H 2 1 Ω .
Applying the ball estimate yields
H 2 4 ≤ C R 2 − n / p H ∥ f ∥ L p ( Ω ) + C R − n / p ∣ Ω ∣ 1 / p H 2 . \frac{H^2}{4}\le CR^{2-n/p}H\|f\|_{L^p(\Omega)}+CR^{-n/p}|\Omega|^{1/p}H^2. 4 H 2 ≤ C R 2 − n / p H ∥ f ∥ L p ( Ω ) + C R − n / p ∣Ω ∣ 1/ p H 2 .
Choose R = K ∣ Ω ∣ 1 / n R = K|\Omega|^{1/n} R = K ∣Ω ∣ 1/ n with K K K sufficiently large to absorb the last term. We obtain H ≤ C ∣ Ω ∣ 2 / n − 1 / p ∥ f ∥ L p ( Ω ) H \le C|\Omega|^{2/n-1/p}\|f\|_{L^p(\Omega)} H ≤ C ∣Ω ∣ 2/ n − 1/ p ∥ f ∥ L p ( Ω ) . Letting δ ↓ 0 \delta\downarrow0 δ ↓ 0 and then ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 proves (37) . For p = ∞ p = \infty p = ∞ , apply the finite-exponent estimate at a fixed admissible exponent and use Hölder’s inequality. □ \square □
Corollary 9.2. Assume the coefficient and exponent hypotheses of Theorem 1.1 or Theorem 1.2. Let Ω \Omega Ω have finite volume and let u ∈ W l o c 2 , p ( Ω ) ∩ C ( Ω ‾ ) u \in W_{\mathrm{loc}}^{2,p}(\Omega) \cap C(\overline{\Omega}) u ∈ W loc 2 , p ( Ω ) ∩ C ( Ω ) be bounded above. Suppose β = b T A − 1 b ∈ L p ( Ω ) \beta= b^{T}A^{-1}b \in L^p(\Omega) β = b T A − 1 b ∈ L p ( Ω ) and f = ( L u ) − ∈ L p ( Ω ) f = (Lu)^{-} \in L^p(\Omega) f = ( Lu ) − ∈ L p ( Ω ) , where L L L is given by (10) . Let C 0 C_0 C 0 be the constant in (37) and set
E = C 0 4 ∣ Ω ∣ 2 / n − 1 / p ∥ β ∥ L p ( Ω ) . E = \frac{C_0}{4}|\Omega|^{2/n-1/p}\|\beta\|_{L^p(\Omega)}. E = 4 C 0 ∣Ω ∣ 2/ n − 1/ p ∥ β ∥ L p ( Ω ) . Then
sup Ω u ≤ sup ∂ Ω u + + C 0 ( e 1 + E − 1 ) ∣ Ω ∣ 2 / n − 1 / p ∥ f ∥ L p ( Ω ) . \sup_{\Omega} u \le\sup_{\partial\Omega} u^{+} + C_0(e^{1+E}-1)|\Omega|^{2/n-1/p}\|f\|_{L^p(\Omega)}. Ω sup u ≤ ∂ Ω sup u + + C 0 ( e 1 + E − 1 ) ∣Ω ∣ 2/ n − 1/ p ∥ f ∥ L p ( Ω ) . On a bounded open set, the same assertion holds with ∣ Ω ∣ 1 / n |\Omega|^{1/n} ∣Ω ∣ 1/ n replaced by diam Ω \operatorname{diam}\Omega diam Ω and C 0 C_0 C 0 by the constant in (2) .
Proof. We use the logarithmic change of variables from [33 ] . Put w = u − sup ∂ Ω u + w = u - \sup_{\partial\Omega}u^{+} w = u − sup ∂ Ω u + and m = sup Ω w m = \sup_{\Omega}w m = sup Ω w . If m ≤ 0 m \le0 m ≤ 0 , the estimate is immediate. For t > 0 t > 0 t > 0 , set
v = log m + t m + t − w . v = \log\frac{m+t}{m+t-w}. v = log m + t − w m + t .
The denominator is at least t t t . Since p > n / 2 p > n/2 p > n /2 , D u ∈ L l o c 2 p Du \in L_{\mathrm{loc}}^{2p} D u ∈ L loc 2 p , so the Sobolev chain rule gives v ∈ W l o c 2 , p ( Ω ) ∩ C ( Ω ‾ ) v \in W_{\mathrm{loc}}^{2,p}(\Omega) \cap C(\overline{\Omega}) v ∈ W loc 2 , p ( Ω ) ∩ C ( Ω ) . Also v ≤ 0 v \le0 v ≤ 0 on the boundary and sup Ω v = log ( 1 + m / t ) \sup_{\Omega}v = \log(1+m/t) sup Ω v = log ( 1 + m / t ) . Using L u ≥ − f Lu \ge-f Lu ≥ − f and completing the square, we obtain
A : D 2 v ≥ − f m + t − w − b ⋅ D v + ( D v ) T A D v ≥ − f t − β 4 . A : D^{2}v \ge-\frac{f}{m+t-w} - b\cdot Dv + (Dv)^{T}ADv \ge-\frac{f}{t} - \frac{\beta}{4}. A : D 2 v ≥ − m + t − w f − b ⋅ D v + ( D v ) T A D v ≥ − t f − 4 β .
The volume estimate applied to v v v therefore gives
log ( 1 + m / t ) ≤ C 0 ∣ Ω ∣ 2 / n − 1 / p t ∥ f ∥ L p ( Ω ) + E . \log(1+m/t) \le\frac{C_0|\Omega|^{2/n-1/p}}{t}\|f\|_{L^p(\Omega)} + E. log ( 1 + m / t ) ≤ t C 0 ∣Ω ∣ 2/ n − 1/ p ∥ f ∥ L p ( Ω ) + E .
Take t = m / ( e 1 + E − 1 ) t = m/(e^{1+E}-1) t = m / ( e 1 + E − 1 ) and rearrange. The diameter version follows by applying (2) to the same function v v v . □ \square □
Radial counterexamples
The radial obstructions go back to Pucci [29 ] ; see also [9 ] . We recall the cutoff constructions that give smooth counterexamples at the critical exponents.
Lemma 10.1. For n ≥ 3 n \ge3 n ≥ 3 , there are nonnegative v j ∈ C ∞ ( B 1 ‾ ) v_j \in C^{\infty}(\overline{B_1}) v j ∈ C ∞ ( B 1 ) , vanishing on ∂ B 1 \partial B_1 ∂ B 1 , such that
∥ v j ∥ L ∞ ( B 1 ) = v j ( 0 ) = 1 , ∥ Δ v j ∥ L n / 2 ( B 1 ) ⟶ 0. \|v_j\|_{L^\infty(B_1)} = v_j(0) = 1,\qquad\|\Delta v_j\|_{L^{n/2}(B_1)} \longrightarrow0. ∥ v j ∥ L ∞ ( B 1 ) = v j ( 0 ) = 1 , ∥Δ v j ∥ L n /2 ( B 1 ) ⟶ 0. Proof. Choose a smooth radial cutoff 0 ≤ χ ε ≤ 1 0 \le\chi_\varepsilon\le1 0 ≤ χ ε ≤ 1 , supported in ( ε , e − 2 ) (\varepsilon,e^{-2}) ( ε , e − 2 ) and equal to one on ( 2 ε , e − 3 ) (2\varepsilon,e^{-3}) ( 2 ε , e − 3 ) , and put
f ε ( r ) = χ ε ( r ) r 2 log ( 1 / r ) . f_\varepsilon(r) = \frac{\chi_\varepsilon(r)}{r^2\log(1/r)}. f ε ( r ) = r 2 log ( 1/ r ) χ ε ( r ) . Its L n / 2 L^{n/2} L n /2 norm is uniformly bounded, because
∫ B 1 f ε n / 2 d x ≤ C n ∫ 0 e − 2 d r r [ log ( 1 / r ) ] n / 2 = C n ∫ 2 ∞ t − n / 2 d t < ∞ . \int_{B_1} f_\varepsilon^{n/2}\,dx \le C_n\int_0^{e^{-2}}\frac{dr}{r[\log(1/r)]^{n/2}} = C_n\int_2^\infty t^{-n/2}\,dt < \infty. ∫ B 1 f ε n /2 d x ≤ C n ∫ 0 e − 2 r [ log ( 1/ r ) ] n /2 d r = C n ∫ 2 ∞ t − n /2 d t < ∞. Let w ε w_\varepsilon w ε be the radial solution of − Δ w ε = f ε -\Delta w_\varepsilon= f_\varepsilon − Δ w ε = f ε with zero boundary values. It is constant near zero, smooth on the closed ball, nonnegative, and decreasing in r r r . Integration of ( r n − 1 w ε ′ ) ′ = − r n − 1 f ε (r^{n-1}w_\varepsilon')' = -r^{n-1}f_\varepsilon ( r n − 1 w ε ′ ) ′ = − r n − 1 f ε gives
w ε ( 0 ) = 1 n − 2 ∫ 0 1 ( r − r n − 1 ) f ε ( r ) d r ⟶ ∞ . w_\varepsilon(0) = \frac{1}{n-2}\int_0^1(r-r^{n-1})f_\varepsilon(r)\,dr \longrightarrow\infty. w ε ( 0 ) = n − 2 1 ∫ 0 1 ( r − r n − 1 ) f ε ( r ) d r ⟶ ∞.
The first term diverges as log log ( 1 / ε ) \log\log(1/\varepsilon) log log ( 1/ ε ) , whereas the second is bounded. Set v ε = w ε / w ε ( 0 ) v_\varepsilon=w_\varepsilon/w_\varepsilon(0) v ε = w ε / w ε ( 0 ) and choose any sequence ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 . □ \square □
We now prove the necessity statement of Theorem 1.1 . If δ ≥ 1 / [ 2 ( n − 1 ) ] \delta\ge1/[2(n-1)] δ ≥ 1/ [ 2 ( n − 1 )] , then P t r = n / 2 P_{\mathrm{tr}}=n/2 P tr = n /2 , and Lemma 10.1 applies with A = I / n A=I/n A = I / n . Suppose 0 < δ < 1 / [ 2 ( n − 1 ) ] 0<\delta<1/[2(n-1)] 0 < δ < 1/ [ 2 ( n − 1 )] , and set
c δ = 1 − ( n − 1 ) δ , μ = ( n − 1 ) δ c δ < 1 , p c = n c δ = n 1 + μ > n 2 . c_\delta=1-(n-1)\delta,\qquad\mu=\frac{(n-1)\delta}{c_\delta}<1,\qquad p_c=nc_\delta=\frac{n}{1+\mu}>\frac{n}{2}. c δ = 1 − ( n − 1 ) δ , μ = c δ ( n − 1 ) δ < 1 , p c = n c δ = 1 + μ n > 2 n .
For x ≠ 0 x\ne0 x = 0 , let
A ( x ) = δ I + ( 1 − n δ ) e r ⊗ e r , e r = x / ∣ x ∣ . A(x)=\delta I+(1-n\delta)e_r\otimes e_r,\qquad e_r=x/|x|. A ( x ) = δ I + ( 1 − n δ ) e r ⊗ e r , e r = x /∣ x ∣.
Its radial eigenvalue is c δ c_\delta c δ , its tangential eigenvalues are δ \delta δ , and its trace is one. For nonnegative smooth radial f f f vanishing near zero, the solution of A : D 2 w = − f A:D^2w=-f A : D 2 w = − f , with w ( 1 ) = 0 w(1)=0 w ( 1 ) = 0 and w w w constant near zero, satisfies
w ′ ( r ) = − c δ − 1 r − μ ∫ 0 r s μ f ( s ) d s , w ( 0 ) = 1 c δ ( 1 − μ ) ∫ 0 1 ( r μ − r ) f ( r ) d r . (38) \begin{aligned}
w'(r)&=-c_\delta^{-1}r^{-\mu}\int_0^r s^\mu f(s)\,ds,\\
w(0)&=\frac{1}{c_\delta(1-\mu)}\int_0^1(r^\mu-r)f(r)\,dr. \tag*{(38)}
\end{aligned} w ′ ( r ) w ( 0 ) = − c δ − 1 r − μ ∫ 0 r s μ f ( s ) d s , = c δ ( 1 − μ ) 1 ∫ 0 1 ( r μ − r ) f ( r ) d r . (38)
The first identity is obtained by multiplying the radial equation by r μ r^\mu r μ . Integrating − w ′ -w' − w ′ and interchanging the two nonnegative integrals gives the second. Choose
f ε ( r ) = χ ε ( r ) r 1 + μ log ( 1 / r ) . f_\varepsilon(r)=\frac{\chi_\varepsilon(r)}{r^{1+\mu}\log(1/r)}. f ε ( r ) = r 1 + μ log ( 1/ r ) χ ε ( r ) .
Since p c ( 1 + μ ) = n p_c(1+\mu)=n p c ( 1 + μ ) = n and p c > 1 p_c>1 p c > 1 , its L p c L^{p_c} L p c norm is uniformly bounded. The r μ f ε r^\mu f_\varepsilon r μ f ε term in (38) diverges as log log ( 1 / ε ) \log\log(1/\varepsilon) log log ( 1/ ε ) ; the r f ε r f_\varepsilon r f ε term remains bounded because μ < 1 \mu<1 μ < 1 . Normalize w ε w_\varepsilon w ε by its maximum w ε ( 0 ) w_\varepsilon(0) w ε ( 0 ) . The resulting functions have maximum one and forcing tending to zero in L p c L^{p_c} L p c .
The coefficients can be made smooth without changing their action on w ε w_\varepsilon w ε . This function is constant for r ≤ ε r\le\varepsilon r ≤ ε . Replace A A A there by a smooth interpolation between I / n I/n I / n for r ≤ ε / 2 r\le\varepsilon/2 r ≤ ε /2 and the displayed radial matrix before r = ε r=\varepsilon r = ε . Convexity preserves the trace and lower bound, and the Hessian vanishes wherever A A A is changed. Finally, on the fixed ball, Hölder’s inequality gives failure at every 1 ≤ p ≤ p c 1\le p\le p_c 1 ≤ p ≤ p c . This proves the lower assertion of Theorem 1.1 .
Counterexamples at and below the spectral exponents
A finite-order laminate on Sym ( d ) \operatorname{Sym}(d) Sym ( d ) is a probability measure obtained from a Dirac mass by finitely many rank-one splittings
δ H ⟼ θ δ H 1 + ( 1 − θ ) δ H 2 , H = θ H 1 + ( 1 − θ ) H 2 , rank ( H 1 − H 2 ) = 1 , \delta_H\longmapsto\theta\delta_{H_1}+(1-\theta)\delta_{H_2},\qquad H=\theta H_1+(1-\theta)H_2,\qquad\operatorname{rank}(H_1-H_2)=1, δ H ⟼ θ δ H 1 + ( 1 − θ ) δ H 2 , H = θ H 1 + ( 1 − θ ) H 2 , rank ( H 1 − H 2 ) = 1 ,
where 0 < θ < 1 0<\theta<1 0 < θ < 1 and all matrices are symmetric. A splitting of an atom of mass ω \omega ω is multiplied by ω \omega ω . We first state the Hessian realization needed to turn such a finite measure into a smooth test function.
Lemma 11.1. Let d ≥ 2 d\ge2 d ≥ 2 , let ν = ∑ i ω i δ H i \nu=\sum_i\omega_i\delta_{H_i} ν = ∑ i ω i δ H i be a finite-order laminate on Sym ( d ) \operatorname{Sym}(d) Sym ( d ) with barycenter H H H , and let G : Sym ( d ) → [ 0 , ∞ ) G:\operatorname{Sym}(d)\to[0,\infty) G : Sym ( d ) → [ 0 , ∞ ) be continuous. For every ε > 0 \varepsilon>0 ε > 0 there is ϕ ∈ C c ∞ ( B 1 d ) \phi\in C_c^\infty(B_1^d) ϕ ∈ C c ∞ ( B 1 d ) such that
∥ ϕ ∥ C 1 < ε , ∣ ∫ B 1 d G ( H + D 2 ϕ ) d x − ∣ B 1 d ∣ ∫ G d ν ∣ < ε . \|\phi\|_{C^1}<\varepsilon,\qquad\left|\int_{B_1^d}G(H+D^2\phi)\,dx-|B_1^d|\int G\,d\nu\right|<\varepsilon. ∥ ϕ ∥ C 1 < ε , ∫ B 1 d G ( H + D 2 ϕ ) d x − ∣ B 1 d ∣ ∫ G d ν < ε .
Proof. Apply the symmetric-matrix part of [2 Proposition 2.3] on an inner ball B r d B_r^d B r d , r < 1 r<1 r < 1 . It realizes the laminate by gradients of Lipschitz maps, with affine boundary data, arbitrarily small uniform distance from that affine map, and uniformly bounded gradient support for the fixed laminate. The realizing map has symmetric gradient. Since the ball is simply connected, it is the gradient of a function w ∈ W 2 , ∞ ( B r d ) w \in W^{2,\infty}(B_r^d) w ∈ W 2 , ∞ ( B r d ) , with D w = H x Dw = Hx D w = H x on the boundary. The Hessian distribution approximates ν \nu ν .
Let q = w − 1 2 x ⋅ H x q = w - \frac{1}{2}x \cdot Hx q = w − 2 1 x ⋅ H x . Then D q = 0 Dq = 0 D q = 0 on the boundary. Since the sphere is connected for d ≥ 2 d \ge2 d ≥ 2 , its boundary value is constant; subtract that constant. Now q q q and D q Dq D q both vanish on the boundary. Smallness of D q Dq D q also makes q q q small in the uniform norm. Extend q q q by zero to B 1 d B_1^d B 1 d and mollify at a scale smaller than the distance between the two spheres. The extension belongs to W 2 , ∞ W^{2,\infty} W 2 , ∞ , and the mollified functions are in C 0 1 ( B 1 d ) C_0^1(B_1^d) C 0 1 ( B 1 d ) . They converge in C 1 C^1 C 1 and in every finite W 2 , s W^{2,s} W 2 , s norm, with a uniform Hessian bound for the fixed realization. Thus continuity of G G G passes its integral to the limit.
Choose r r r close enough to one that the omitted shell has negligible contribution, including its constant Hessian H H H . Then choose the realization error small, and finally the mollification scale small. These successive choices prove both stated inequalities. □
We also need smooth coefficients near the minimizing Pucci selection. For a symmetric matrix H H H , define
A τ ( H ) = I + ρ − 1 2 [ I − H ( H 2 + τ 2 I ) − 1 / 2 ] , τ > 0. (39) A_\tau(H) = I + \frac{\rho-1}{2}\left[I-H(H^2+\tau^2I)^{-1/2}\right], \qquad\tau>0. \tag*{(39)} A τ ( H ) = I + 2 ρ − 1 [ I − H ( H 2 + τ 2 I ) − 1/2 ] , τ > 0. (39)
This is smooth in H H H , satisfies I ≤ A τ ( H ) ≤ ρ I I \le A_\tau(H) \le\rho I I ≤ A τ ( H ) ≤ ρ I , and
0 ≤ A τ ( H ) : H − M 1 , ρ − ( H ) ≤ d ( ρ − 1 ) 2 τ . (40) 0 \le A_\tau(H):H-\mathcal{M}_{1,\rho}^{-}(H) \le\frac{d(\rho-1)}{2}\tau. \tag*{(40)} 0 ≤ A τ ( H ) : H − M 1 , ρ − ( H ) ≤ 2 d ( ρ − 1 ) τ . (40)
Diagonalization reduces these claims to − 1 < t / t 2 + τ 2 < 1 -1 < t/\sqrt{t^2+\tau^2} < 1 − 1 < t / t 2 + τ 2 < 1 and 0 ≤ ∣ t ∣ − t 2 / t 2 + τ 2 ≤ τ 0 \le|t|-t^2/\sqrt{t^2+\tau^2} \le\tau 0 ≤ ∣ t ∣ − t 2 / t 2 + τ 2 ≤ τ . Thus small L p L^p L p norm of ( M 1 , ρ − ( − D 2 u ) ) + (\mathcal{M}_{1,\rho}^{-}(-D^2u))_+ ( M 1 , ρ − ( − D 2 u ) ) + also gives small forcing for smooth coefficients, by choosing τ \tau τ small.
Fix d ≥ 2 d \ge2 d ≥ 2 and 1 ≤ m < d 1 \le m < d 1 ≤ m < d . Write κ ( a ) \kappa(a) κ ( a ) for the smallest eigenvalue of K d , m ( a ) K_{d,m}(a) K d , m ( a ) from (32) . For h m = 0 h_m=0 h m = 0 , this matrix acts by
( K d , m ( a ) h ) j = d − j 1 + a ( h j − h j + 1 ) + j a 1 + a ( h j − h j − 1 ) , 0 ≤ j < m , (K_{d,m}(a)h)_j = \frac{d-j}{1+a}(h_j-h_{j+1})+\frac{ja}{1+a}(h_j-h_{j-1}), \qquad0 \le j < m, ( K d , m ( a ) h ) j = 1 + a d − j ( h j − h j + 1 ) + 1 + a j a ( h j − h j − 1 ) , 0 ≤ j < m ,
where the second term is absent at j = 0 j=0 j = 0 .
Lemma 11.2. The eigenvalue κ ( a ) \kappa(a) κ ( a ) is positive and simple, and κ ′ ( a ) < 0 \kappa'(a)<0 κ ′ ( a ) < 0 . Its positive right eigenvector, normalized by h 0 = 1 h_0=1 h 0 = 1 , satisfies
1 = h 0 > h 1 > ⋯ > h m − 1 > h m = 0. 1=h_0>h_1>\cdots>h_{m-1}>h_m=0. 1 = h 0 > h 1 > ⋯ > h m − 1 > h m = 0.
Proof. Positivity, simplicity, and the positive eigenvector follow from Lemma 8.1. Let π j = ( d j ) a − j \pi_j=\binom{d}{j}a^{-j} π j = ( j d ) a − j . Multiply the eigenvalue equations by π j \pi_j π j and sum from 0 0 0 to j j j . Detailed balance π j λ j = π j + 1 μ j + 1 \pi_j\lambda_j=\pi_{j+1}\mu_{j+1} π j λ j = π j + 1 μ j + 1 cancels the interior terms and gives
π j λ j ( h j − h j + 1 ) = κ ∑ i = 0 j π i h i > 0. \pi_j\lambda_j(h_j-h_{j+1})=\kappa\sum_{i=0}^{j}\pi_i h_i>0. π j λ j ( h j − h j + 1 ) = κ i = 0 ∑ j π i h i > 0.
This proves strict decrease of h h h . A positive left eigenvector is ℓ j = π j h j \ell_j=\pi_jh_j ℓ j = π j h j . The coefficients depend smoothly on a a a and the eigenvalue is simple, so differentiation of the eigenvalue equation yields
κ ′ ( a ) = ℓ T K d , m ′ ( a ) h ℓ T h . \kappa'(a)=\frac{\ell^T K'_{d,m}(a)h}{\ell^T h}. κ ′ ( a ) = ℓ T h ℓ T K d , m ′ ( a ) h .
The derivative applied to h h h , with h h h held fixed, has entries
( K d , m ′ ( a ) h ) j = − d − j ( 1 + a ) 2 ( h j − h j + 1 ) + j ( 1 + a ) 2 ( h j − h j − 1 ) < 0. (K'_{d,m}(a)h)_j=-\frac{d-j}{(1+a)^2}(h_j-h_{j+1})+\frac{j}{(1+a)^2}(h_j-h_{j-1})<0. ( K d , m ′ ( a ) h ) j = − ( 1 + a ) 2 d − j ( h j − h j + 1 ) + ( 1 + a ) 2 j ( h j − h j − 1 ) < 0.
The first term is strictly negative and the second is nonpositive. The positive left eigenvector therefore gives κ ′ ( a ) < 0 \kappa'(a)<0 κ ′ ( a ) < 0 . □
Proposition 11.3. Let d ≥ 2 d \ge2 d ≥ 2 , 1 ≤ m < d 1 \le m < d 1 ≤ m < d , and ρ ≥ 1 \rho\ge1 ρ ≥ 1 . If 1 ≤ p < κ d , m ( ρ ) 1 \le p < \kappa_{d,m}(\rho) 1 ≤ p < κ d , m ( ρ ) , or if p = κ d , m ( ρ ) > 1 p = \kappa_{d,m}(\rho) > 1 p = κ d , m ( ρ ) > 1 , then there are u j ∈ C ∞ ( B 1 d ) u_j \in C^\infty(B^d_1) u j ∈ C ∞ ( B 1 d ) , zero on the boundary, and smooth symmetric fields I ≤ A j ≤ ρ I I \le A_j \le\rho I I ≤ A j ≤ ρ I such that
∥ u j ∥ L ∞ ( B 1 d ) ≤ 3 4 , sup B 1 d u j ≥ 1 4 , ∥ ( A j : D 2 u j ) − ∥ L p ( B 1 d ) ⟶ 0. \|u_j\|_{L^\infty(B^d_1)} \le\frac{3}{4}, \qquad\sup_{B^d_1} u_j \ge\frac{1}{4}, \qquad\|(A_j : D^2u_j)^-\|_{L^p(B^d_1)} \longrightarrow0. ∥ u j ∥ L ∞ ( B 1 d ) ≤ 4 3 , B 1 d sup u j ≥ 4 1 , ∥ ( A j : D 2 u j ) − ∥ L p ( B 1 d ) ⟶ 0.
Proof. Set a 0 = ( d − m ) / ( m ρ ) a_0 = (d-m)/(m\rho) a 0 = ( d − m ) / ( m ρ ) and a = a 0 − η a = a_0-\eta a = a 0 − η , where
0 ≤ η ≤ η 0 < a 0 , η 0 < d m ρ ( m + 1 ) . (41) 0 \le\eta\le\eta_0 < a_0, \qquad\eta_0 < \frac{d}{m\rho(m+1)}. \tag*{(41)} 0 ≤ η ≤ η 0 < a 0 , η 0 < m ρ ( m + 1 ) d . (41)
At scale t > 0 t>0 t > 0 , use diagonal matrices with entries in { t , − a t } \{t,-at\} { t , − a t } . If a matrix has j j j negative entries, then
M 1 , ρ − ( H ) = t ( d − j − ρ a j ) = t [ d ( 1 − j m ) + ρ η j ] . (42) \mathcal{M}_{1,\rho}^{-}(H) = t(d-j-\rho aj) = t\left[d\left(1-\frac{j}{m}\right)+\rho\eta j\right]. \tag*{(42)} M 1 , ρ − ( H ) = t ( d − j − ρ aj ) = t [ d ( 1 − m j ) + ρ η j ] . (42)
We stop a branch at the end of a round when it has at least m m m negative entries. If it has exactly m m m , its Pucci value is m ρ η t m\rho\eta t m ρ η t . If it has more than m m m , its Pucci value is nonpositive. To see the latter, the most restrictive condition is at j = m + 1 j=m+1 j = m + 1 , and it is exactly ensured by (41) ; the quantity ( j − m ) / j (j-m)/j ( j − m ) / j increases with j j j .
Choose s ∈ ( 0 , 1 ) s\in(0,1) s ∈ ( 0 , 1 ) . In one round, replace each of the d d d coordinates, one at a time, by either t / s t/s t / s or − a t / s -at/s − a t / s . The probability of the positive choice is
u s = s + a 1 + a u_s=\frac{s+a}{1+a} u s = 1 + a s + a
for an old positive entry,
z s = a ( 1 − s ) 1 + a z_s=\frac{a(1-s)}{1+a} z s = 1 + a a ( 1 − s )
for an old negative entry.
For example, u s ( t / s ) + ( 1 − u s ) ( − a t / s ) = t u_s(t/s)+(1-u_s)(-at/s)=t u s ( t / s ) + ( 1 − u s ) ( − a t / s ) = t ; the other formula gives − a t -at − a t . Thus each replacement preserves the parent matrix in expectation. Only one diagonal entry changes, so every replacement is a symmetric rank-one split. Stopping is decided only after all d d d coordinates have been replaced.
Let T = T ( s , a ) T=T(s,a) T = T ( s , a ) be the transition matrix for the counts 0 , … , m − 1 0,\ldots,m-1 0 , … , m − 1 that have not been stopped. Independence of the coordinate choices gives
T i j = ∑ r = max { 0 , j − d + i } min { i , j } ( i r ) ( 1 − z s ) r z s i − r ( d − i j − r ) ( 1 − u s ) j − r u s d − i − j + r . (43) T_{ij}=\sum_{r=\max\{0,j-d+i\}}^{\min\{i,j\}}
\binom{i}{r}(1-z_s)^r z_s^{i-r}
\binom{d-i}{j-r}(1-u_s)^{j-r}u_s^{d-i-j+r}. \tag*{(43)} T ij = r = m a x { 0 , j − d + i } ∑ m i n { i , j } ( r i ) ( 1 − z s ) r z s i − r ( j − r d − i ) ( 1 − u s ) j − r u s d − i − j + r . (43)
Here r r r old negative entries remain negative, and j − r j-r j − r old positive entries become negative. Write ϵ = 1 − s \epsilon=1-s ϵ = 1 − s . A positive entry changes sign with probability ϵ / ( 1 + a ) \epsilon/(1+a) ϵ / ( 1 + a ) , while a negative entry changes sign with probability a ϵ / ( 1 + a ) a\epsilon/(1+a) a ϵ / ( 1 + a ) . Events with at least two sign changes have probability O ( ϵ 2 ) O(\epsilon^2) O ( ϵ 2 ) . Thus (43) gives, uniformly for a ∈ [ a 0 − η 0 , a 0 ] a\in[a_0-\eta_0,a_0] a ∈ [ a 0 − η 0 , a 0 ] ,
T ( s , a ) = I − ϵ K d , m ( a ) + O ( ϵ 2 ) . (44) T(s,a)=I-\epsilon K_{d,m}(a)+O(\epsilon^2). \tag*{(44)} T ( s , a ) = I − ϵ K d , m ( a ) + O ( ϵ 2 ) . (44)
The loss of transitions to count m m m is why the final diagonal of K K K includes the absorption rate.
Suppose κ ( a ) > p \kappa(a)>p κ ( a ) > p and put Δ = κ ( a ) − p \Delta=\kappa(a)-p Δ = κ ( a ) − p . The normalized positive eigenvector h ( a ) h(a) h ( a ) is continuous in a a a . After reducing η 0 \eta_0 η 0 , there is h ∗ > 0 h_* > 0 h ∗ > 0 such that
h ∗ ≤ h j ( a ) ≤ 1 ( 0 ≤ j < m ) . h_* \le h_j(a) \le1 \qquad(0 \le j < m). h ∗ ≤ h j ( a ) ≤ 1 ( 0 ≤ j < m ) .
For each fixed a a a , choose ϵ > 0 \epsilon>0 ϵ > 0 small enough that the error in (44) , applied to h h h , is at most ( Δ / 2 ) ϵ h (\Delta/2)\epsilon h ( Δ/2 ) ϵ h componentwise. We may also require
T h ≤ c h , c = 1 − ( p + Δ / 2 ) ϵ > 0. Th \le ch, \qquad c=1-(p+\Delta/2)\epsilon>0. T h ≤ c h , c = 1 − ( p + Δ/2 ) ϵ > 0.
Convexity of s ↦ s p s\mapsto s^p s ↦ s p for p ≥ 1 p\ge1 p ≥ 1 gives
s p − c ≥ Δ ϵ 2 > 0. s^p-c\ge\frac{\Delta\epsilon}{2}>0. s p − c ≥ 2 Δ ϵ > 0.
Fix such an s s s for this a a a , and set r = c / s p < 1 r=c/s^p<1 r = c / s p < 1 . Starting at I d I_d I d , perform N N N rounds on branches not yet stopped. The resulting finite laminate ν η , N \nu_{\eta,N} ν η , N has barycenter I d I_d I d . Put G ( H ) = ( M 1 , ρ ( H + ) ) p G(H)=(\mathcal{M}_{1,\rho}(H^+))^p G ( H ) = ( M 1 , ρ ( H + ) ) p . Let b j b_j b j be the probability that a round starting with j < m j<m j < m negative entries ends with exactly m m m . This requires at least one sign change, and each coordinate changes sign with probability at most ϵ \epsilon ϵ . Therefore
0 ≤ b j ≤ d ϵ , b ≤ ( d / h ∗ ) ϵ h . 0\le b_j\le d\epsilon,\qquad b\le(d/h_*)\epsilon h. 0 ≤ b j ≤ d ϵ , b ≤ ( d / h ∗ ) ϵ h .
By (42) , a branch first stopped in round q q q has scale s − q s^{-q} s − q and positive Pucci value m ρ η s − q m\rho\eta s^{-q} m ρ η s − q when its count is exactly m m m . Writing e 0 e_0 e 0 for the first coordinate vector, its contribution summed over all such branches is
E s t o p p e d = ( m ρ η ) p ∑ q = 1 N s − q p e 0 T T q − 1 b ≤ C η p ϵ s − p ∑ q = 0 N − 1 r q ≤ C η p ϵ s p − c ≤ C η p Δ . \begin{aligned}
E_{\mathrm{stopped}}=(m\rho\eta)^p\sum_{q=1}^{N}s^{-qp}e_0^{T}T^{q-1}b \\
&\le C\eta^p\epsilon s^{-p}\sum_{q=0}^{N-1}r^q\le\frac{C\eta^p\epsilon}{s^p-c}\le\frac{C\eta^p}{\Delta}.
\end{aligned} E stopped = ( m ρ η ) p q = 1 ∑ N s − q p e 0 T T q − 1 b ≤ C η p ϵ s − p q = 0 ∑ N − 1 r q ≤ s p − c C η p ϵ ≤ Δ C η p .
Branches stopped with more than m m m negative entries have zero energy. On a surviving branch the scale-one positive Pucci value is at most d d d . Since 1 ≤ h / h ∗ 1\le h/h_* 1 ≤ h / h ∗ and T N h ≤ c N h T^Nh\le c^Nh T N h ≤ c N h , the surviving energy is at most C r N Cr^N C r N . Thus
∫ G d ν η , N ≤ C η p κ ( a ) − p + C r N . (45) \int G\,d\nu_{\eta,N}\le\frac{C\eta^p}{\kappa(a)-p}+Cr^N. \tag*{(45)} ∫ G d ν η , N ≤ κ ( a ) − p C η p + C r N . (45)
The constants are uniform in η \eta η , although s s s and r r r may depend on η \eta η .
If p < κ ( a 0 ) p<\kappa(a_0) p < κ ( a 0 ) , take η = 0 \eta=0 η = 0 , choose s s s once, and let N → ∞ N\to\infty N → ∞ . The first term in (45) vanishes. If p = κ ( a 0 ) > 1 p=\kappa(a_0)>1 p = κ ( a 0 ) > 1 , Lemma 11.2 gives κ ( a 0 − η ) − p ≥ c 0 η \kappa(a_0-\eta)-p\ge c_0\eta κ ( a 0 − η ) − p ≥ c 0 η for small positive η \eta η . Consequently
∫ G d ν η , N ≤ C η p − 1 + C r N . \int G\,d\nu_{\eta,N}\le C\eta^{p-1}+Cr^N. ∫ G d ν η , N ≤ C η p − 1 + C r N .
First let η \eta η tend to zero. For each such η \eta η , choose s s s as above and then a finite N N N large enough to make r N r^N r N tend to zero. Here η p − 1 → 0 \eta^{p-1}\to0 η p − 1 → 0 because p > 1 p>1 p > 1 .
Apply Lemma 11.1 separately to the resulting finite laminates, with energy error tending to zero and ∥ ϕ j ∥ C 0 < 1 / 4 \|\phi_j\|_{C^0}<1/4 ∥ ϕ j ∥ C 0 < 1/4 . Define
u j ( x ) = 1 − ∣ x ∣ 2 2 − ϕ j ( x ) . u_j(x)=\frac{1-|x|^2}{2}-\phi_j(x). u j ( x ) = 2 1 − ∣ x ∣ 2 − ϕ j ( x ) .
These functions vanish on the boundary, have absolute value at most 3 / 4 3/4 3/4 , and satisfy u j ( 0 ) > 1 / 4 u_j(0)>1/4 u j ( 0 ) > 1/4 . Moreover, − D 2 u j = I d + D 2 ϕ j -D^2u_j=I_d+D^2\phi_j − D 2 u j = I d + D 2 ϕ j , so G ( − D 2 u j ) → 0 G(-D^2u_j)\to0 G ( − D 2 u j ) → 0 . Take A j = A τ j ( − D 2 u j ) A_j=A_{\tau_j}(-D^2u_j) A j = A τ j ( − D 2 u j ) from (39) , with τ j ↓ 0 \tau_j\downarrow0 τ j ↓ 0 . By (40) ,
( A j : D 2 u j ) − = ( A j : ( − D 2 u j ) ) + ≤ ( M 1 , ρ ( − D 2 u j ) ) + + d ( ρ − 1 ) 2 τ j . (A_j:D^2u_j)^-=(A_j:(-D^2u_j))^+\le(\mathcal{M}_{1,\rho}(-D^2u_j))^++\frac{d(\rho-1)}{2}\tau_j. ( A j : D 2 u j ) − = ( A j : ( − D 2 u j ) ) + ≤ ( M 1 , ρ ( − D 2 u j ) ) + + 2 d ( ρ − 1 ) τ j .
This proves the claimed forcing convergence with smooth coefficients. □
Passage to higher dimensions
We must account for each lower-dimensional term in (8) . A direct product cutoff would differentiate the original test functions and give uncontrolled terms. The following convex truncation avoids that problem and also preserves the endpoint.
Lemma 12.1. Let 1 ≤ d < n 1\le d<n 1 ≤ d < n , 1 ≤ p < ∞ 1\le p<\infty 1 ≤ p < ∞ , and ρ ≥ 1 \rho\ge1 ρ ≥ 1 . Suppose u j ∈ C ∞ ( B 1 d ‾ ) u_j\in C^\infty(\overline{B_1^d}) u j ∈ C ∞ ( B 1 d ) vanish on the boundary and there are measurable fields I d ≤ A j ≤ ρ I d I_d\le A_j\le\rho I_d I d ≤ A j ≤ ρ I d and constants C , c > 0 C,c>0 C , c > 0 such that
∥ u j ∥ ∞ ≤ C , sup u j ≥ c , ∥ ( A j : D 2 u j ) − ∥ L p ( B 1 d ) ⟶ 0. \|u_j\|_\infty\le C,\qquad\sup u_j\ge c,\qquad\|(A_j:D^2u_j)^-\|_{L^p(B_1^d)}\longrightarrow0. ∥ u j ∥ ∞ ≤ C , sup u j ≥ c , ∥ ( A j : D 2 u j ) − ∥ L p ( B 1 d ) ⟶ 0.
Then there are V j ∈ C ∞ ( B 1 n ) V_j \in C^\infty(B_1^n) V j ∈ C ∞ ( B 1 n ) and smooth symmetric fields I n ≤ B j ≤ ρ I n I_n \le B_j \le\rho I_n I n ≤ B j ≤ ρ I n such that
∥ V j ∥ ∞ ≤ C , sup V j ≥ c / 2 , ∥ ( B j : D 2 V j ) − ∥ L p ( B 1 n ) ⟶ 0. \lVert V_j\rVert_\infty\le C,\qquad\sup V_j \ge c/2,\qquad\lVert(B_j:D^2V_j)^-\rVert_{L^p(B_1^n)} \longrightarrow0. ∥ V j ∥ ∞ ≤ C , sup V j ≥ c /2 , ∥( B j : D 2 V j ) − ∥ L p ( B 1 n ) ⟶ 0.
Proof. Put q = n − d q=n-d q = n − d and fix 0 < η < c / 4 0<\eta<c/4 0 < η < c /4 . Choose a smooth convex function χ : R → [ 0 , ∞ ) \chi:\mathbb{R}\to[0,\infty) χ : R → [ 0 , ∞ ) with
χ = 0 on ( − ∞ , 0 ] , 0 ≤ χ ′ ≤ 1 , χ ′ ′ ≥ 0 , ( t − η ) + ≤ χ ( t ) ≤ t + . \chi=0\text{ on }(-\infty,0],\qquad0\le\chi'\le1,\qquad\chi''\ge0,\qquad(t-\eta)_+\le\chi(t)\le t_+. χ = 0 on ( − ∞ , 0 ] , 0 ≤ χ ′ ≤ 1 , χ ′′ ≥ 0 , ( t − η ) + ≤ χ ( t ) ≤ t + .
For 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 , set
U j , δ ( x , z ) = u j ( x ) − δ ∣ z ∣ 2 − η , v j , δ = χ ( U j , δ ) , ( x , z ) ∈ B 1 d × R q . U_{j,\delta}(x,z)=u_j(x)-\delta\lvert z\rvert^2-\eta,\qquad v_{j,\delta}=\chi(U_{j,\delta}),\qquad(x,z)\in B_1^d\times\mathbb{R}^q. U j , δ ( x , z ) = u j ( x ) − δ ∣ z ∣ 2 − η , v j , δ = χ ( U j , δ ) , ( x , z ) ∈ B 1 d × R q .
For each j j j , continuity and the zero boundary values give a collar of ∂ B 1 d \partial B_1^d ∂ B 1 d where u j < η u_j<\eta u j < η . Hence extension by zero makes v j , δ v_{j,\delta} v j , δ smooth with compact support in R n \mathbb{R}^n R n , and
supp v j , δ ⊂ B 1 d × B C / δ q , ∥ v j , δ ∥ ∞ ≤ C , sup v j , δ ≥ c − 2 η > c / 2. \operatorname{supp}v_{j,\delta}\subset B_1^d\times B_{\sqrt{C/\delta}}^q,\qquad\lVert v_{j,\delta}\rVert_\infty\le C,\qquad\sup v_{j,\delta}\ge c-2\eta>c/2. supp v j , δ ⊂ B 1 d × B C / δ q , ∥ v j , δ ∥ ∞ ≤ C , sup v j , δ ≥ c − 2 η > c /2.
Extend A j A_j A j by I d I_d I d outside B 1 d B_1^d B 1 d and put A ~ j = diag ( A j , I q ) \widetilde A_j=\operatorname{diag}(A_j,I_q) A j = diag ( A j , I q ) . Writing f j = ( A j : D 2 u j ) − f_j=(A_j:D^2u_j)^- f j = ( A j : D 2 u j ) − , the chain rule and χ ′ ′ ≥ 0 \chi''\ge0 χ ′′ ≥ 0 give, on the support,
A ~ j : D 2 v j , δ = χ ′ ( U j , δ ) ( A j : D 2 u j − 2 q δ ) + χ ′ ′ ( U j , δ ) ( D U j , δ ) T A ~ j D U j , δ ≥ − f j − 2 q δ . \widetilde A_j:D^2v_{j,\delta}
=\chi'(U_{j,\delta})(A_j:D^2u_j-2q\delta)+\chi''(U_{j,\delta})(DU_{j,\delta})^T\widetilde A_jDU_{j,\delta}
\ge-f_j-2q\delta. A j : D 2 v j , δ = χ ′ ( U j , δ ) ( A j : D 2 u j − 2 q δ ) + χ ′′ ( U j , δ ) ( D U j , δ ) T A j D U j , δ ≥ − f j − 2 q δ .
The support bound therefore yields
∥ ( A ~ j : D 2 v j , δ ) − ∥ L p ( R n ) ≤ C 1 ( δ − q / ( 2 p ) ∥ f j ∥ L p ( B 1 d ) + δ 1 − q / ( 2 p ) ) . \lVert(\widetilde A_j:D^2v_{j,\delta})^-\rVert_{L^p(\mathbb{R}^n)}
\le C_1\left(\delta^{-q/(2p)}\lVert f_j\rVert_{L^p(B_1^d)}+\delta^{1-q/(2p)}\right). ∥( A j : D 2 v j , δ ) − ∥ L p ( R n ) ≤ C 1 ( δ − q / ( 2 p ) ∥ f j ∥ L p ( B 1 d ) + δ 1 − q / ( 2 p ) ) .
Set R δ = 2 1 + C / δ R_\delta=2\sqrt{1+C/\delta} R δ = 2 1 + C / δ and define V j , δ ( y ) = v j , δ ( R δ y ) V_{j,\delta}(y)=v_{j,\delta}(R_\delta y) V j , δ ( y ) = v j , δ ( R δ y ) and B j , δ ( y ) = A ~ j ( R δ y ) B_{j,\delta}(y)=\widetilde A_j(R_\delta y) B j , δ ( y ) = A j ( R δ y ) . The support lies in B 1 / 2 n B_{1/2}^n B 1/2 n . The Hessian gains a factor R δ 2 R_\delta^2 R δ 2 , while the L p L^p L p norm gains R δ − n / p R_\delta^{-n/p} R δ − n / p . Since R δ R_\delta R δ is comparable to δ − 1 / 2 \delta^{-1/2} δ − 1/2 , we obtain
∥ ( B j , δ : D 2 V j , δ ) − ∥ L p ( B 1 / 2 n ) ≤ C 2 ( δ − 1 + d / ( 2 p ) ∥ f j ∥ L p ( B 1 d ) + δ d / ( 2 p ) ) . (46) \lVert(B_{j,\delta}:D^2V_{j,\delta})^-\rVert_{L^p(B_{1/2}^n)}
\le C_2\left(\delta^{-1+d/(2p)}\lVert f_j\rVert_{L^p(B_1^d)}+\delta^{d/(2p)}\right). \tag*{(46)} ∥( B j , δ : D 2 V j , δ ) − ∥ L p ( B 1/2 n ) ≤ C 2 ( δ − 1 + d / ( 2 p ) ∥ f j ∥ L p ( B 1 d ) + δ d / ( 2 p ) ) . (46)
To make (46) tend to zero, choose any sequence δ ℓ ↓ 0 \delta_\ell\downarrow0 δ ℓ ↓ 0 . For each ℓ \ell ℓ , choose j ℓ j_\ell j ℓ large enough that the first term on the right tends to zero. The second term tends to zero without a restriction on p p p . This proves the assertions with measurable coefficients.
Finally let H = − D 2 V j ℓ , δ ℓ H=-D^2V_{j_\ell,\delta_\ell} H = − D 2 V j ℓ , δ ℓ . Since M 1 , ρ − ( H ) ≤ B j ℓ , δ ℓ : H M_{1,\rho}^-(H)\le B_{j_\ell,\delta_\ell}:H M 1 , ρ − ( H ) ≤ B j ℓ , δ ℓ : H , the smooth selection (39) has forcing no greater than the previous forcing plus n ( ρ − 1 ) τ / 2 n(\rho-1)\tau/2 n ( ρ − 1 ) τ /2 . Choose τ ↓ 0 \tau\downarrow0 τ ↓ 0 along the sequence to obtain the required smooth fields. □
We can now complete Theorem 1.2. For 1 ≤ p ≤ n / 2 1\le p\le n/2 1 ≤ p ≤ n /2 , Lemma 10.1 gives counterexamples with constant coefficients. If n / 2 < p < P n ( ρ ) n/2<p<P_n(\rho) n /2 < p < P n ( ρ ) , some pair ( d , m ) (d,m) ( d , m ) in the finite maximum (8) satisfies p < κ d , m ( ρ ) p<\kappa_{d,m}(\rho) p < κ d , m ( ρ ) . Proposition 11.3 gives the counterexample in dimension d d d ; when d < n d<n d < n , apply Lemma 12.1. At p = P n ( ρ ) > n / 2 p=P_n(\rho)>n/2 p = P n ( ρ ) > n /2 , a maximizing pair satisfies p = κ d , m ( ρ ) > 1 p=\kappa_{d,m}(\rho)>1 p = κ d , m ( ρ ) > 1 , so the endpoint part of the same proposition applies. The case P n ( ρ ) = n / 2 P_n(\rho)=n/2 P n ( ρ ) = n /2 has already been treated. Multiplying the coefficient fields by λ \lambda λ restores λ I ≤ A j ≤ Λ I \lambda I\le A_j\le\Lambda I λ I ≤ A j ≤ Λ I and only multiplies their forcing by λ \lambda λ . The necessity assertion is proved.
Shibing Chen
Email address: chenshibing1982@hotmail.com
Neil S. Trudinger, Mathematical Sciences Institute, The Australian National University, Canberra ACT 0200, Australia.
Email address: Neil.Trudinger@anu.edu.au
Xu-Jia Wang, Institute for Theoretical Sciences, Westlake University, Hangzhou, 310030, China.
Email address: wangxujia@westlake.edu.cn